Dissociating Controllable and Uncontrollable Effects of Affective Stimuli on Attitudes and Consumption

Dissociating Controllable and Uncontrollable Effects of Affective Stimuli on Attitudes and... Abstract This research studies a fundamental and seemingly straightforward question: Can basic advertising elements, such as the presence of attractive imagery, have uncontrollable effects on consumers’ attitudes and consumption decisions? Answering this question is methodologically challenging, because the presence of an uncontrollable process can be masked by a simultaneously operating controllable process. We argue first that existing methods conflate the contribution of both processes and are therefore unable to measure the presence of an uncontrollable process reliably. To solve the conundrum, we present a novel application of processing tree modeling. Evaluative conditioning is employed as a paradigm to study the influence of affective visual stimuli on attitudes and behavior. Across six experiments, we demonstrate the validity of the model parameters estimating controllable and uncontrollable processes. As predicted, the parameter estimate of the controllable process is susceptible to cognitive resources and levels of motivation to exert control. The parameter estimate of the uncontrollable process appears unaffected by these factors. We also demonstrate the external validity of our findings and their relevance to stimuli and instructions typical for consumer research. Finally, we find that controllable and uncontrollable processes are both predictive of product choice and consumption. We discuss implications for consumer protection. attitudes, evaluative conditioning, evaluative learning, controllability, automaticity, multinomial processing tree models The evaluative conditioning (EC) procedure is an experimental analog to the acquisition of consumer attitudes in everyday life. In a typical EC procedure, a conditioned stimulus (CS; e.g., a brand name) repeatedly co-occurs in close spatiotemporal proximity with a valenced unconditioned stimulus (US; e.g., a beautiful image or a celebrity endorser). As a result, evaluations of the CS typically change in the direction of the valence of the US—a phenomenon known as the EC effect. This basic procedure is considered a key representation of evaluative learning in general (for reviews, see De Houwer, Thomas, and Baeyens 2001; Hofmann et al. 2010). EC also serves as the prototypical paradigm to study advertising effects on brand attitudes (Allen and Madden 1985; Bierley, McSweeney, and Vannieuwkerk 1985; Gibson 2008; Gorn 1982; Janiszewski and Warlop 1993; Shimp, Stuart, and Engle 1991; Stuart, Shimp, and Engle 1987; Sweldens, Van Osselaer, and Janiszewski 2010). In the present research, we propose and validate a novel methodological approach to assess whether and to what degree EC procedures can have uncontrollable effects on attitudes and consumption. As we will outline, we study controllability first, because it can play a crucial role in a theoretical debate that has been going on for decades: Are the processes underlying EC effects characterized by features of automaticity? The second reason is substantive in nature—such uncontrollable effects would have important implications for consumer welfare and protection. INTRODUCTION In the introduction to this article, we first highlight the theoretical relevance of studying controllability. Next, we illuminate the methodological approach we propose, introducing a multinomial processing tree model to separate controllable from uncontrollable processes in EC. We then propose and present a series of studies to test and validate this model. We return to the substantive implications for consumer protection in the General Discussion. Automaticity, Controllability, and EC The possible contribution of automatically operating learning processes has a long and contentious history in EC. Consistent with Williams and Poehlman’s (2017) recent treatise on consciousness, much of this debate has focused on whether EC effects can be established without awareness of the contingencies between the stimuli (Sweldens, Corneille, and Yzerbyt 2014). However, the concept of automaticity is defined by several features that do not necessarily co-occur (Bargh 1994; Williams and Poehlman 2017). Instead, cognitive processes can be characterized by automaticity in varying configurations. Therefore, it is unwarranted to infer the presence or absence of other automaticity features from evidence regarding awareness (Fiedler and Hütter 2014; Moors and De Houwer 2006; Sweldens, Tuk, and Hütter 2017). Notably, research investigating automaticity features other than awareness often indicates that EC would not be automatic in these respects. For example, three recent articles investigating EC’s efficiency found that EC effects do not occur when participants are under cognitive load (Dedonder et al. 2010; Mierop, Hütter, and Corneille 2017; Pleyers et al. 2009). Similarly, research indicates that EC effects are influenced by intentionality or participants’ processing goals during the EC procedure (Corneille et al. 2009; Fiedler and Unkelbach 2011). Studying the controllability of the learning process in EC has the potential to advance the debate on this front significantly. Current State of the Art: Explicit Evaluations Can, but Implicit Evaluations Cannot, Be Controlled We know of only two previous studies investigating the controllability of EC effects. Gawronski, Balas, and Creighton (2014) instructed participants to prevent or promote the effect of the US pairings on their CS evaluations in the conditioning procedure. The authors observed a dissociation between an indirect measure of evaluation (an evaluative priming task) and a direct self-report measure. While the indirect measure showed standard, unmodulated EC effects in all conditions, the direct measure appeared subject to expressive control. The same pattern of results was obtained in subsequent work comparing different strategies of exerting control. Irrespective of the strategy participants employed (e.g., suppressing emotional reactions to the US or reappraising the valence of the US), it appeared that indirect measures of evaluation were not subject to control, but deliberate responses on direct measures appeared controllable (Gawronski, Mitchell, and Balas 2015). In summary, this research suggests that, while the associative structure might change (as revealed in the indirect measures of evaluation), consumers retain the ability to control their explicit evaluations. Why We Investigate the Uncontrollability of Direct Evaluations There are two important reasons why the current research focuses on direct rather than indirect measures: a methodological and a substantive one. First, whereas indirect measures of evaluation are often good reflections of associative structure, they may not give a complete picture of what people have learned. This issue is especially acute in contexts where negations come into play. For instance, Deutsch, Gawronski, and Strack (2006) presented their participants with affirmed or negated primes in an evaluative priming task. In indirect measures of evaluation, attitudes toward primes are inferred from participants’ response times in classifying subsequent target stimuli. The presentation of the prime “pleasure” facilitated responses for positive targets and inhibited responses for negative targets. However, Deutsch and colleagues (2006) found the same pattern of response facilitation and inhibition when the prime “no pleasure” was presented. Hence, a negated representation still improved reaction times for targets of the opposite valence. These results illustrate why it is problematic to rely on indirect measures to study the controllability of EC. Consider the case of a controllability experiment in which participants aim to reverse or negate the effects of US valence on CS evaluations. If EC effects would be completely controllable, participants might encode CSs that were shown with positive USs as associated with “not positive” instead. However, when this CS is later encountered as a prime in an evaluative priming task, its “not positive” association can still facilitate response times to positive targets and thereby indicate a regular EC effect, even though full control was exerted over learning and participants actually hold an association of opposite valence. Hence, reaction-time-based evaluative measures contain the risk of considerably overestimating the contribution of uncontrollable learning, which should be avoided in a research context where the very existence of a phenomenon (in this case, uncontrollability) is to be demonstrated. Second, from a substantive viewpoint it is similarly important to study direct evaluations. As indicated above, according to the current state of the literature, EC can have uncontrollable effects on indirect measures of evaluation, implying that conditioning procedures can change concepts’ associative structure uncontrollably (Gawronski et al. 2014; Gawronski et al. 2015). At the same time, the current evidence indicates that control can still be exerted at the expression stage, when consumers validate the output of the associative system (i.e., in their direct evaluations). Substantively, one could argue that it does not matter at which stage control is being exerted. As long as consumers are successful at negating the effects of EC (e.g., by invalidating the output of the associative system), their autonomy has not been compromised as they retain control over their deliberate behavior. If, however, EC would have uncontrollable effects on direct evaluations, this would be more problematic from a consumer protection viewpoint. After all, this would indicate that consumers’ control strategies failed at both the encoding and the validation stages, opening the door for uncontrollable effects on deliberate behavior as well. Why We Need Better Measurement Methods The main methodological challenge in studying the uncontrollability of EC stems from the difficulty of distinguishing the contribution of multiple processes in any measurement outcome. The presence of uncontrollable effects needs to be detected in addition to the controllable effects of EC. This problem originates in the fact that no measure of evaluation is process-pure: both direct and indirect measures of evaluation can be reflective of both explicitly and implicitly acquired evaluations (De Houwer 2006; De Houwer et al. 2009). To illustrate this problem in the current context, consider the following, hypothetical example of a researcher who aims to study the controllability of EC. The researcher designs an EC experiment featuring a conditioning procedure in which several CSs are paired with positive USs and other CSs are paired with negative USs. To study the controllability of the process, the researcher implements three experimental conditions. In the control condition, participants receive no special instructions as they go through the conditioning procedure and subsequently provide their evaluative ratings of the CS on a measure of evaluation (whether the measure is direct or indirect does not matter for this example). In the prevent condition, participants are asked to prevent any influence the conditioning procedure might have on their evaluative ratings of the CSs. In the reverse condition, participants are asked to reverse any influence the conditioning procedure might have—that is, to evaluate CSs paired with positive images more negatively than those paired with negative images. Consider the hypothetical pattern of results displayed in figure 1. FIGURE 1 View largeDownload slide HYPOTHETICAL RESULTS OF AN EXPERIMENT ON CONTROLLABILITY IN EC NOTE.— CSs– = Negatively paired conditioned stimuli, CSs+ = positively paired conditioned stimuli. FIGURE 1 View largeDownload slide HYPOTHETICAL RESULTS OF AN EXPERIMENT ON CONTROLLABILITY IN EC NOTE.— CSs– = Negatively paired conditioned stimuli, CSs+ = positively paired conditioned stimuli. At first glance, the researcher could be tempted to conclude that the EC effect is completely under participants’ control. After all, consistent with the hypothesis that the effect is controllable, the researcher observes a standard EC effect in the control condition, no EC effect in the prevent condition, and a reversed EC effect in the reverse condition. However, a more careful look at the results reveals the limitations of the method. First, the null result in the prevent condition is, like most null results, difficult to interpret. Even if EC had an uncontrollable effect on attitudes, participants in this condition could still easily comply with the task instructions to prevent the influence of the conditioning procedure on their evaluative ratings by employing several response strategies (e.g., by clicking the midpoint of the scale in the case of direct measures, or by responding randomly in the case of either direct or indirect measures of evaluation). Second, the reversed EC effect in the reverse condition is also ambiguous. In this condition, the controllable process may stand in opposition to an uncontrollable one. In such cases, one should account for the possibility that the uncontrollable process might generate a smaller (but non-negligible) effect than the controllable process. In an experimental setting with many CSs, control might, for example, fail for a (statistically significant) minority of CSs, whereas for the majority of CSs control succeeds and a reversed EC effect is observed. This would be consistent with work studying other features of automaticity in EC. For example, research investigating EC’s dependence on awareness of the CS-US contingency found a much larger proportion of CSs showing a conditioning effect characterized by contingency memory than CSs showing a conditioning effect without contingency memory (Hütter et al. 2012; Hütter and Sweldens 2013). The researcher’s problem is that an aggregated measure of evaluation does not distinguish between a situation in which participants can reverse evaluations for some CSs but not for others, and a competing possibility in which participants are in fact able to reverse the EC effect for all CSs but form less extreme evaluations. Whereas the former possibility would be consistent with the existence of an uncontrollable process in EC, the latter possibility would imply that EC is fully controllable. Overcoming the problems caused by aggregation requires the use of the processing tree framework to distinguish the contributions of different processes on the level of trials in a single measure. THE PROCESSING TREE FRAMEWORK Processing tree models are a class of statistical models that allow for separating and quantifying the effects of processes underlying performance in a single task (Batchelder and Riefer 1999). In the present research, we develop a multinomial processing tree (MPT) model to measure the contribution of controllable and uncontrollable processes to the formation of evaluations in a direct measure. Processing tree models have in common that latent cognitive processes are related to observable responses. The contribution of each process is estimated in terms of a parameter that expresses the probability of that process operating. Processing tree models can be applied to categorical data (e.g., “pleasant” and “unpleasant” responses) with the restriction that the database needs to be rich enough to allow for the separation of multiple processes. Therefore, in many cases processing trees require the use of different experimental conditions that lead to distinct relations of the latent cognitive processes to the observable responses. This framework has been applied in a variety of contexts (Hütter and Klauer 2016; Payne and Bishara 2009). The present work examines whether this framework is suitable for investigating the controllability of EC effects. That is, we develop and validate an MPT model to separate controllable and uncontrollable attitude acquisition processes in an EC paradigm. To dissociate the processes contributing (i.e., the latent constructs) to EC effects (i.e., the observable responses), we create a standard and a reversal version of the conditioning phase following the process dissociation logic introduced by Jacoby (1991). In the standard condition, participants are told that they can apply the valence of the USs to form an evaluation of the CSs. In this condition, both controllable and uncontrollable processes lead to regular EC effects. In the reversal condition, the instructions state that the US valence should be opposite to the one presented for participants to form accurate evaluations of the CS. To the extent that US valence has uncontrollable effects on CS attitudes, the CS will still acquire the unmodified US valence. Hence, in the reversal condition, controllable and uncontrollable processes have opposite effects on CS evaluations. These predictions can be visualized in terms of the processing tree model depicted in figure 2. The model relates contributions of controllable valence transfer (estimated by the c-parameter) and uncontrollable valence transfer (the u-parameter) to dichotomous evaluative responses while controlling for unsystematic processes (i.e., that do not vary systematically with US valence and instructions; the r-parameter). FIGURE 2 View largeDownload slide PROCESSING TREE MODEL OF POSITIVE (+) AND NEGATIVE (–) EVALUATIONS FORMED IN THE CONDITIONING PHASE IN THE STANDARD AND REVERSAL CONDITIONS FOR A POSITIVELY (CS+) AND NEGATIVELY (CS–) PAIRED CS NOTE.—The rectangles on the left denote the stimuli, while the rectangles on the right denote the responses. The branches of the processing tree represent the combination of cognitive processes postulated by the model. c = probability of acquiring an evaluation via controllable processes; u = probability of acquiring an evaluation via uncontrollable processes given the absence of controllable processes; r = positive response tendency. FIGURE 2 View largeDownload slide PROCESSING TREE MODEL OF POSITIVE (+) AND NEGATIVE (–) EVALUATIONS FORMED IN THE CONDITIONING PHASE IN THE STANDARD AND REVERSAL CONDITIONS FOR A POSITIVELY (CS+) AND NEGATIVELY (CS–) PAIRED CS NOTE.—The rectangles on the left denote the stimuli, while the rectangles on the right denote the responses. The branches of the processing tree represent the combination of cognitive processes postulated by the model. c = probability of acquiring an evaluation via controllable processes; u = probability of acquiring an evaluation via uncontrollable processes given the absence of controllable processes; r = positive response tendency. Several explanatory notes are in order. First, the main theoretical interest of our research lies not in demonstrating instances of failure of control per se (1 – c), but in whether there is an uncontrollable process (u) contributing to evaluations if control fails (this does not necessarily follow from the former). To elaborate, the model perfectly allows for trials or instances where people do not exert control or follow instructions—that is, where they do not apply US valence to the CS evaluation as instructed (1 – c). There are plenty of reasons why people might not exert control that are not due to an uncontrollable effect of the affective stimuli. For instance, they might not be motivated to follow instructions or they might not be paying attention to the stimuli, which would all lower the c-parameter (and hence increase 1 – c). It is important to realize, however, that none of those circumstances on their own would necessarily increase the u-parameter, because in all those circumstances it remains possible that no valence is transferred from the US to the CS. The u-parameter is derived only from those cases (across conditions) where US valence is applied irrespective of the instructions given to the participant. Therefore, the u-parameter strictly measures the uncontrolled application of US valence. Consequently, in a second step it still needs to be demonstrated that such instances of uncontrolled US valence transfer are characterized by features of automaticity, before one regards the u-parameter as a reflection of uncontrollable (i.e., automatic) valence transfer. Second, it is important to note that because the MPT model is applied to dichotomous response options, it is not necessary to imagine a stimulus or generate an affective reaction that is equally strong as the actual US presented for control to be registered by the model. The quality (i.e., positive or negative valence), but not the quantity (i.e., intensity), is critical in producing the c-parameter. Third, the MPT model is specifically designed to test for the presence of uncontrollable processes over and above controllable ones, without making any assumptions about the sequence of processes operating during encoding. That is, the model should neither imply that uncontrollable processes would operate only when control fails nor deny the possibility that uncontrollable processes take precedence over controllable ones during encoding. The model constitutes a mere measurement model that assesses whether there is a contribution of uncontrollable processes (u-parameter) to attitude acquisition after controllable processes (c-parameter) and unsystematic processes (r-parameter) have been partialed out. Finally, the explicit nature of the instructions has several consequences that are important for theorizing on automatic attitude acquisition. Specifically, the instructions used in this controllability setting are quite explicit about the CS-US pairings and their expected consequences for CS attitudes. Hence, awareness levels can be expected to be high in the present research. High levels of awareness, however, do not guarantee high levels of controllability. To the contrary, participants may be aware of the stimuli, their pairing, and their evaluative reaction to the US, but not be able to reverse the effect of US valence on the evaluation of the CS. Experiments in the domain of thought control illustrate this fact. Even though individuals are fully aware of unwanted thoughts entering consciousness, these thoughts are difficult to control. In fact, attempting to exert control increases the probability of unwanted thoughts occurring in consciousness (Wegner 1994; Whitmer and Banich 2007). Validating the Model and Its Parameters in a Series of Studies Developing and applying an MPT model in a new research context does not by itself guarantee the reliability and validity of the parameters obtained (Hütter and Klauer 2016). We follow a multistudy approach to demonstrate the method’s ability to distinguish controllable from uncontrollable processes in EC, reflected by its c- and u-parameters. This requires demonstrating several desired characteristics of the model and parameter estimates. The first indicator is model fit; that is, in all studies we assess whether the model we developed constitutes a suitable description of the data obtained. Fit statistics assess the degree to which the predicted response frequencies deviate from the observed response frequencies on the categorical evaluative measure across experimental conditions. The asymptotically χ2-distributed log-likelihood statistic G² is used to assess the goodness of fit of MPT models (Hu and Batchelder 1994). A good fit (i.e., low deviation) would be expected when the model can accommodate the observed category frequencies. The goodness-of-fit statistic also aids hypothesis testing by comparing different specifications of the MPT model. Specifically, the statistic ΔG² assesses the change in fit when parameters are restricted to a specific value (i.e., zero) or equal to another. Setting parameters to zero and observing the effect on model fit is a way to test whether a parameter is significantly greater than zero (i.e., it explains a substantial proportion of the data) and must be retained in the model. Conducting this test on the uncontrollability (u) parameter is of primary theoretical interest in the present line of research. Model fit alone, however, does not guarantee that the model parameters measure the constructs they are intended to measure (Hütter and Klauer 2016). Hence, the second major challenge is to demonstrate the construct validity of the model parameters, relating to the crucial question of whether the c- and u-parameters indeed capture processes that differ in their degree of automaticity (Bargh 1994). A particularly difficult question is whether the process captured by the u-parameter does indeed reflect uncontrollable effects of US valence on CS attitudes (i.e., it demonstrates automaticity). This is not guaranteed by the MPT parameters per se: a significant u-parameter merely indicates the presence of a statistically significant number of CSs where control was not applied, yet a (non-reversed) EC effect was acquired. This does not necessarily imply automaticity; perhaps control could have been exerted if participants had been more motivated or had more opportunity for processing. To demonstrate the automaticity of a process, one needs to demonstrate that it bypasses manipulations of processing capacity or motivation to control (Fazio 1990). In contrast, the controllability of a process as represented by the c-parameter should depend on whether participants are motivated to exert control and whether they have the cognitive resources to do so. Studies 2 and 3 therefore test the differential susceptibility of the c- and u-parameters to manipulations of cognitive capacity and motivation, establishing their construct validity. We will also investigate the external validity of the method—that is, whether the method and findings are relevant to more than one specific paradigm. We thus investigate the current questions in paradigms that employ faces (studies 1a, 1b, 2, 3) and consumer brands (studies 4, 5) as CSs. Moreover, we also investigate the relevance of our approach for the broader EC literature. Because the main theoretical interest applies to the question of whether visual affective stimuli can have uncontrollable effects, we normally implore participants to exert control to the best of their ability. This is, however, atypical for EC research, where the purpose of the stimulus pairings is generally concealed. Therefore, in study 4 we will create a context that allows us to assess the relevance of our conclusions for more typical EC paradigms. Equally important is to demonstrate the predictive validity of the model parameter estimates—that is, to demonstrate that the model’s parameter estimates can predict scores on other criterion variables. A first important question is whether the MPT parameter estimates correlate as predicted with independently assessed attitude measures. Specifically, the model’s c-parameter, which measures the controllable influence of US valence, should correlate positively with the EC effect in the standard condition, but correlate inversely in a condition in which participants aim to reverse the influence of US valence. Conversely, the model’s u-parameter, which measures the uncontrollable influence of US valence, should correlate positively with the EC effect in both the standard and reversal conditions, as this process should not be affected by instructions or participants’ efforts to exert control. One key advantage of within-participant designs is that parameter estimates are obtained at the individual level (between-participants designs provide only aggregated parameter estimates per condition). Therefore, predictive validity of this kind can be assessed in studies featuring within-participant designs (studies 1b, 3, and 5). The second aspect of predictive validity, and from a substantive point of view arguably the most important one, is whether the model parameter estimates also predict actual product choice and consumption. Investigating the predictive validity of the parameters for consumers’ brand choice and consumption is the main goal of study 5. STUDY 1A: A BETWEEN-PARTICIPANTS TEST OF THE MPT MODEL By comparing responses in the standard and reversal conditions, study 1a investigates whether there is a significant number of CSs for which participants are unable to reverse the US valence when required. The standard and reversal instruction conditions are implemented between-participants, implying that participants need to reverse either no or all US valence information. Method Participants In this first study, 37 students (25 female) of different majors at the University of Heidelberg took part (Mage = 22.35, SDage = 4.81). Participants chose between receiving course credit or €3.00 (approximately US$4.00) as a reward for this study, which took about 25 minutes. Design The experiment employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: preratings vs. postratings) × 2 (instruction condition: standard vs. reversal) mixed design with repeated measures on the first two factors. Materials A set of 102 black-and-white pictures of human faces was used as the CS repertory (Hütter et al. 2012). For each participant, CSs were selected from that pool based on an initial evaluative rating. Fifty unambiguously pleasant and 50 unambiguously unpleasant pictures from the International Affective Picture System (IAPS; Lang, Bradley, and Cuthbert 2008) served as USs. Pleasant and unpleasant pictures differed in valence, t(98) = 60.77, p < .001. They also differed in dominance, t(98) = 19.66, p < .001 (unpleasant pictures were relatively more dominant), but not in arousal, t(98) = –.02, p = .98. Procedure Participants first rated the valence of the 102 portraits on a continuous scale with the endpoints “very unpleasant” (–100) and “very pleasant” (100). For each participant, the 24 portraits with the most neutral rating were selected as CSs. In the subsequent conditioning phase, the CSs were paired with the USs. Each CS was randomly assigned six different USs. Each CS was shown once with each assigned US in a simultaneous presentation. Pairings were presented for 3,000 milliseconds with an intertrial interval of 200 milliseconds. In the standard condition, participants were told that the US valence is informative regarding the person (CS). In the reversal condition, instructions also stated that US valence should be opposite from what is presented. Hence, when an unpleasant picture is presented it should actually be pleasant, and vice versa (precise instructions are in appendix A). After the conditioning phase, participants indicated for every CS whether their evaluation of it was “pleasant” or “unpleasant.” This is the key dichotomous measure to which the MPT model can be applied. Additionally, we obtained continuous evaluative ratings of the 24 CSs a second time (on the same scale as the preratings). After the completion of the experiment, participants were thanked and dismissed. Results Evaluative Ratings The unstandardized continuous evaluative ratings are presented for the two instruction conditions in figure 3. They were analyzed with a 2 (US valence) × 2 (time of evaluative rating) × 2 (instruction condition) mixed ANOVA. EC effects in this design are revealed by the interaction between the US valence and time of evaluative ratings factors, as the EC procedure should affect CS evaluations dependent on US valence. There was no overall EC effect, as the two-way interaction between US valence and time of measurement was not significant (F(1, 35) = 2.35, p = .13, ηp2 = .06). Instead, the EC effect was moderated by instruction condition (F(1, 35) = 21.86, p < .001, ηp2 = .38). Additionally, the overall non-significant main effect of US valence (F(1, 35) = 1.68, p = .20, ηp2 = .05) was dependent on instruction condition (F(1, 35) = 17.84, p < .001, ηp2 = .34). The remaining effects were non-significant (smallest p = .47). FIGURE 3 View largeDownload slide EVALUATIVE RATINGS IN STUDY 1A BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. FIGURE 3 View largeDownload slide EVALUATIVE RATINGS IN STUDY 1A BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. The interaction of US valence and time of measurement was assessed separately for the instruction conditions, revealing a significant EC effect in the standard condition (F(1, 35) = 18.76, p < .001, ηp2 = .56) and a significant reversed EC effect (F(1, 35) = 5.18, p = .03, ηp2 = .20) in the reversal condition. The size of the normal EC effect in the standard condition is almost three times as large as the size of the reversed EC effect in the reversal condition, as revealed by the partial η². An auxiliary test in which the sign of the EC effect is set positive in both standard and reversal conditions showed, however, that this size difference is not significant in this study (F(1, 35) = 2.35, p = .13, ηp2 = .06). MPT Model The frequency data from the dichotomous evaluation task were analyzed with the software HMMTree (Stahl and Klauer 2007; see appendix B for the frequency data of all experiments). The model fit the empirical data well (G²(1) = 0.19, p = .66), implying that the observed frequencies did not differ significantly from the frequencies predicted by the model. The estimate of the c-parameter was c = .28 (95% confidence interval (CI) [.21, .34]), thus reflecting 28% of the CS evaluations. The u-parameter’s estimate was u = .14, 95% CI [.06, .23], accounting for 10% of the pairings (i.e., the product of the converse probability of the c-parameter and the probability expressed by the u-parameter). Setting this parameter to zero led to a significant decrease in model fit (ΔG²(1) = 10.26, p < .001), indicating a significant contribution of the uncontrollable process to attitude formation. The parameter representing response tendencies was r = .44, 95% CI [.39, .50], indicating a slight tendency to respond “unpleasant” as r < .50; see figure 4 for the parameter estimates and their standard errors. FIGURE 4 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 1A NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 4 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 1A NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. Discussion The results clearly illustrate why the use of a MPT methodology is expedient to distinguish controllable from uncontrollable processes, as the results obtained with a classic measure of evaluation would be very hard to interpret unambiguously. On the one hand, we observe a reversed EC effect in the reversal condition, which could be interpreted as indicative of the controllable nature of EC. On the other hand, in the standard condition both the controllable and uncontrollable process lead to a regular EC effect, while in the reversal condition they have opposite effects on evaluations. Therefore, one could argue that the smaller size of the reversed EC effect in the reversal condition constitutes evidence for the uncontrollable nature of EC (although the difference in size was not significant in this study, it is significant in studies 2, 4, and 5). However, significant or not, evidence of this kind could not be regarded as conclusive, as alternative processes might lead to the same outcome (e.g., participants might be able to execute the reversals at will, but form less extreme evaluations for CSs when they perform the reversal). Crucially, the MPT model allows for dissociating and quantifying the contributions of controllable (c-parameter) and uncontrollable processes (u-parameter). Whereas a substantial subset of CSs acquired an evaluation in line with the instructed US valence as evidenced by the c-parameter, the experiment also yielded evidence for uncontrollable EC effects reflected by a significant u-parameter that captures evaluations in line with US valence irrespective of instructions. Thus, both types of processes concurrently contribute to evaluations. STUDY 1B: A WITHIN-PARTICIPANT TEST OF THE MPT MODEL In the next study, we implemented a within-participant operationalization of standard and reversal conditions. Obtaining similar results with a different operationalization of the method would enhance faith in the validity of the method. Specifically, it is possible that the c- and u-parameters could reflect standard and reversed responding at test rather than learning. A precondition for such response strategies would be that participants remember the US valence or acquired an EC effect (and respond accordingly in the standard condition, and give reversed responses in the reversal condition). As participants completed only one instruction condition in the between-participants design, it was theoretically possible to apply the instructions at test rather than during learning. By introducing a within-participant design that interspersed the CSs from both instruction conditions at test, we reduced this possibility, as participants would need to memorize additional information on the instructions pertaining to each CS to report the correct valence at test. If a reversal response set was responsible for the parameter structure obtained, the c-parameter should be smaller in the within- than in the between-participants design. An additional advantage of using a within-participant operationalization is that it provides the opportunity to estimate the c- and u-parameters for each participant individually. This, in turn, allows us to assess whether the parameter estimates predict participants’ ratings on an independent measure of evaluation. Observing that the parameter estimates co-vary as predicted with other measures of evaluation would enhance faith in the validity of the method by providing evidence of predictive validity. Finally, obtaining convergence of the parameter estimates across between- and within-participant operationalizations assuages concerns regarding violations of some assumptions underlying the MPT methodology (e.g., aggregation independence and parameter homogeneity, which are assumed when aggregating across participants; Curran and Hintzman 1995; Jacoby and Shrout 1997). Method Participants Twenty-three students of different majors at the University of Heidelberg took part in the study (Mage = 21.60, SDage = 2.80). We excluded three participants due to anomalous data in the dichotomous evaluation task, with one participant always reversing responses (even in the standard condition when no reversals were required), one responding “unpleasant” in 88% of the cases, and one responding “pleasant” in 92% of the cases. They were identified as outliers (Tukey 1977). The final sample contained 14 females and six males (as there are no between-participants manipulations, we considered this sample size sufficient to test the validity of the method; see our note on sample sizes and power preceding the General Discussion). Participants received course credit or €4.00 (approximately US$5.00). Design The study employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: preratings vs. postratings) × 2 (instruction condition: standard vs. reversal) design with repeated measures on all three factors. Sequence of instructions (standard-first vs. reversal-first) was counterbalanced between participants. Materials and Procedure The materials and procedure followed study 1a with one critical exception. Participants watched two conditioning blocks, one under standard and one under reversal instructions. Hence, the 24 CSs were split between the standard and reversal blocks, which contained 12 CSs each to be presented six times with a US. Dependent measures were administered only after both conditioning blocks were completed with CSs from both blocks commingled in random order in each measure. Results Evaluative Ratings Evaluative ratings were submitted to a 2 (US valence) × 2 (time of measurement) × 2 (instruction condition) × 2 (sequence of instructions) mixed ANOVA. The EC effect was significant (F(1, 18) = 5.17, p = .04, ηp2 = .22) and moderated by instruction condition (F(1, 18) = 20.74, p < .001, ηp2 = .54). Whereas there was no overall main effect of US valence (F(1, 18) = 0.39, p = .54, ηp2 = .02), the effect of US valence was dependent on instruction condition (F(1, 18) = 23.41, p < .001, ηp2 = .57). None of the other effects was significant (smallest p = .15). The EC effects were assessed separately for the two instruction conditions. The standard condition resulted in a significant EC effect (F(1, 19) = 41.77, p < .001, ηp2 = .68). The reversed EC effect in the reversal condition did not reach the conventional level of significance (F(1, 19) = 3.99, p = .06, ηp2 = .17). The partial η² reveal that the size of the EC effect in the standard condition is more than three times the size of the reversed EC effect in the reversal condition. An auxiliary test demonstrated this size difference was significant (F(1, 19) = 5.64, p = .03, ηp2 = .23). The evaluative ratings are displayed in figure 5. FIGURE 5 View largeDownload slide EVALUATIVE RATINGS IN STUDY 1B BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.—The error bars represent standard errors. FIGURE 5 View largeDownload slide EVALUATIVE RATINGS IN STUDY 1B BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.—The error bars represent standard errors. MPT Model The initial model containing one c-, one u-, and one r-parameter per sequence condition described the data well (G²(2) = 1.68, p = .43). The initial parameter estimates were c = .37, 95% CI [.25, .49], u = .15, 95% CI [.00, .35], and r = .46, 95% CI [.35, .58] for the standard-first condition. The parameter estimates for the reversal-first condition were c = .40, 95% CI [.29, .51], u = .19, 95% CI [.01, .37], and r = .52, 95% CI [.41, .63]. The u-parameter differed from zero in both the standard-first (ΔG²(1) = 2.24, p = .07) and reversal-first (ΔG²(1) = 4.01, p = .02) conditions. The c-parameters (ΔG²(1) = 0.15, p = .70), u-parameters (ΔG²(1) = 0.08, p = .78), and r-parameters (ΔG²(1) = 0.47, p = .49) can be set equal across sequence conditions without loss in model fit. The resulting parameter estimates were c = .39, 95% CI [.31, .47], u = .17, 95% CI [.04, .30] (which is significantly larger than zero, ΔG²(1) = 6.19, p < .01), and r = .49, 95% CI [.41, .57] (see figure 6). The final model fit the data well (G²(5) = 2.38, p = .79; frequency data are in appendix B). FIGURE 6 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 1B NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 6 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 1B NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. The MPT data can be modeled jointly for the between- (study 1a) and the within-participant version (study 1b). The initial model with one c-, one u-, and one r-parameter for each experiment showed a good fit (G²(2) = 1.87, p = .39). Setting the c-parameters equal across experiments impaired the model’s fit. In contrast to the reversal-at-response explanation, the c-parameter for the within-participant variant, cwithin = .39, 95% CI [.31, .47], was significantly larger than the one for the between-participants experiment, cbetween = .28, 95% CI [.21, .34], ΔG²(1) = 4.57, p = .03. The power analysis presented before the General Discussion indicates, however, that this difference between designs was relatively small. The u-parameters can be set equal without a loss in model fit, ΔG²(1) = 0.12, p = .73. The resulting u-parameter with an estimate of u = .15, 95% CI [.08, .23] was significantly larger than zero, ΔG²(1) = 16.70, p < .001. The estimates of the r-parameter also did not differ between conditions, ΔG²(1) = 0.98, p = .32, resulting in a combined estimate of r = .44, 95% CI [.42, .50]. The final model fit the data well, G²(4) = 2.97, p = .56. Predictive Value of Individual Parameter Estimates The next analysis investigates whether the parameter estimates obtained from the dichotomous evaluation task (in which participants rated CSs as pleasant or unpleasant) are predictive of CS evaluations on a separate, continuous measure. Standardized parameter values for the controllable and uncontrollable components were estimated per participant. We calculated EC effects from the continuous evaluative ratings by subtracting the mean evaluation of the negatively paired CSs from the mean evaluation of the CSs that have been paired positively. We specified a Hierarchical Linear Model (HLM) predicting the EC effects from the instruction condition, the c- and u-parameters, and their interaction with the instruction condition (the model is hierarchical, as the evaluative ratings for both instruction conditions are nested within participants). The main effect of instruction condition was significant, as the EC effect is positive in the standard condition, but negative in the reversal condition (B = –41.42, SE = 7.27, t(16) = –5.69, p < .001). The main effect of the c-parameter was not significant (B = –8.65, SE = 10.19, t(16) = –0.85, p = .41), but was moderated by instruction condition (B = –76.40, SE = 26.32, t(16) = –2.90, p = .01). As expected, the c-parameter shows a (marginally significant) positive relation to the EC effect in the standard condition (B = 29.55, SE = 16.64, t(16) = 1.78, p = .09) and a negative relation in the reversal condition (B = –46.85, SE = 16.65, t(16) = –2.81, p = .01). The u-parameter, on the other hand, is a universally positive predictor of the EC effect (B = 24.65, SE = 8.73, t(16) = 2.82, p = .01), irrespective of the instructions participants received (B = 23.12, SE = 22.55, t(16) = 1.03, p = .32). Discussion This study replicates study 1a in a within-participant design, increasing our confidence in the validity of the parameter estimates in three different ways. First, the fact that parameters were equal in size across sequence conditions alleviates potential concerns that switching between tasks would bias the parameters (Rogers and Monsell 1995). Second, the fact that the c-parameter was not impaired (but rather enhanced) in the within-participant variant suggests that participants executed the reversals during learning rather than at test. Third, the within-participant design enables estimating for every participant the contribution of controllable and uncontrollable processes from responses in the first, dichotomous evaluation task. Both parameter estimates predict the EC effect on a second, continuous measure of evaluation in the expected directions, demonstrating their predictive validity. Specifically, whereas the effect of the c-parameter depends on the instruction condition, the u-parameter has a universally positive influence on the EC effect, independent of instructions. The fact that the u-parameter successfully predicts evaluations over and above the c-parameter on a continuous direct measure of evaluation implies that the results obtained by direct measures are a combination of both controllable and uncontrollable processes. The (smaller) reversed EC effect observed on the direct measure in the reversal condition is thus composed of a mix of controllable, reversed EC effects and uncontrollable, standard EC effects. STUDY 2: MANIPULATION OF PROCESSING CAPACITY A key challenge of the current research is to establish the construct validity of the MPT parameters. We claim that the model’s c-parameter reflects the contribution of a controllable cognitive process, whereas the u-parameter reflects the contribution of an uncontrollable process. One can provide construct validity for MPT model parameters by demonstrating that the parameter values respond as predicted to experimental manipulations known to affect nonautomatic and automatic processes differently (Hütter et al. 2012). One primary variable of interest is working memory capacity. The controllability of a process should hinge on working memory capacity, as deliberative effort is required and negation operations are known to require extensive cognitive resources (Bargh 1994; Deutsch et al. 2009; Fazio 1990). Hence, our trust in the validity of the c-paramater as capturing a controllable cognitive process would be enhanced if its effect is reduced when cognitive capacity is impaired. In addition, manipulating cognitive capacity can provide evidence for the thesis that the u-parameter reflects an uncontrollable process that is independent of cognitive resources. This thesis would be supported if the u-parameter were equally large under conditions with full cognitive capacity as under conditions where cognitive capacity is constrained. Method Participants Fifty-nine students (49 female) of different majors at the University of Heidelberg took part in this experiment (Mage = 22.29, SDage = 4.50). Participants chose between course credit and €4.00 (approximately US$5.00). Design We employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: pre- vs. postratings) × 2 (instruction condition: standard vs. reversal) × 2 (cognitive load: low vs. high) mixed design with repeated measures on the first two factors. Materials and Procedure The materials were identical to those in studies 1a and 1b. We employed several procedural changes in order to establish a low- and a high-load condition. The conditioning phase was segmented into six blocks. In the high-load condition, a four-digit number was presented for six seconds before each block, which participants were asked to remember. In the low-load condition, no number was presented and participants saw the message “Please wait. To be continued shortly.” After each block, participants were asked to type in a four-digit number (the memorized one in the high-load condition; any number in the low-load condition). We decided to limit the number of digits to four mainly because the task of controlling evaluative learning is already demanding and previous research showed that high levels of cognitive and attentional demand can completely eliminate EC effects (Dedonder et al. 2010; Pleyers et al. 2009). Note that the goal of this manipulation was not to completely absorb processing resources but to demonstrate a relative dependence on processing resources of one process that is not present in the other. The sequence of CSs within each block was randomized. Results Evaluative Ratings Evaluative ratings, illustrated in figure 7, were submitted to a 2 (US valence) × 2 (time of evaluative rating) × 2 (instruction condition) × 2 (cognitive load) mixed ANOVA. The overall EC effect (i.e., the US valence × time of measurement interaction) was significant (F(1, 55) = 4.54, p = .04, ηp2 =.08), involving also a significant main effect of US valence (F(1, 55) = 6.62, p = .01, ηp2 = .11). The EC effect was significantly moderated by instruction condition (F(1, 55) = 12.82, p < .001, ηp2 = .19) as was the main effect of US valence (F(1, 55) = 10.62, p < .01, ηp2 = .16). Whereas the EC effect was not moderated by cognitive load (F(1, 55) = 0.00, p = 1, ηp2 = 0.00), its interaction with instruction condition was significantly moderated by load (F(1, 55) = 5.73, p = .02, ηp2 = 0.02). There was a marginally significant main effect of instruction condition (F(1, 55) = 3.48, p = .07, ηp2 = .09). All other effects were nonsignificant (largest F = 2.48, smallest p = .12). FIGURE 7 View largeDownload slide EVALUATIVE RATINGS IN STUDY 2 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND COGNITIVE LOAD CONDITION NOTE.— The error bars represent standard errors. FIGURE 7 View largeDownload slide EVALUATIVE RATINGS IN STUDY 2 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND COGNITIVE LOAD CONDITION NOTE.— The error bars represent standard errors. To decompose the four-way interaction, we calculated separate 2 (US valence) × 2 (time of evaluative rating) ANOVAs for each cell of the between-participant design (load × instruction condition). Without cognitive load, there was a regular EC effect in the standard condition (F(1, 14) = 39.80, p < .001, ηp2 = .74), while the reversed EC effect in the reversal condition was not significant (F(1, 13) = 2.39, p = .15, ηp2 = .16). Under cognitive load, the standard condition also resulted in a reliable EC effect (F(1, 14) = 5.76, p = .03, ηp2 = .29), while the pairings had no effect in the reversal condition (F(1, 14) = 0.14, p = .72, ηp2 = .01). In the low-load condition, the size of the EC effect in the standard condition was more than twice the size of the reversed EC effect in the reversal condition, as revealed by the partial η². However, this difference failed to reach the conventional significance level in this study (F(1, 27) = 2.41, p = .13, ηp2 = .08). MPT Model An initial model was fit to the frequency data (in appendix B) containing one c-, one u-, and one r-parameter per load condition. Initial model fit was good (G²(2) = 2.62, p = .27). Estimates of the parameters in the low-load condition were c = .28, 95% CI [.21, .35], u = .13, 95% CI [.03, .23], and r = .48, 95% CI [.42, .54]. In the high-load condition, they were c = .07, 95% CI [.00, .14], u = .07, 95% CI [.00, .15], and r = .50, 95% CI [.45, .54]. See figure 8 for the parameter estimates and their standard errors. The individual u-parameters differed significantly from zero in both the low-load (ΔG² (1) = 6.74, p < .01) and the high-load conditions (ΔG² (1) = 3.22, p = .04). While the u-parameters (ΔG²(1) = 0.82, p = .37) and the r-parameters (ΔG²(1) = 0.23, p = .63) did not differ between the cognitive load conditions, the estimate for the c-parameter was significantly larger in the low-load condition than in the high-load condition (ΔG²(1) = 15.62, p < .001). The resulting model containing the four parameters clow = .28, 95% CI [.21, .35], chigh = .07, 95% CI [.00, .15], u = .09, 95% CI [.03, .16], and r = .49, 95% CI [.46, .52], fit the data well (G²(4) = 3.67, p = .45). The global u-parameter was significantly larger than zero (ΔG² (1) = 9.07, p < .01). FIGURE 8 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 2 NOTE.— c denotes the controllable parameter with separate estimates for the high-load and the low-load conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 8 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 2 NOTE.— c denotes the controllable parameter with separate estimates for the high-load and the low-load conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. Discussion A cognitive load manipulation was introduced to establish construct validity of the c- and u-parameters as capturing an effortful, controllable process versus an efficient, uncontrollable process, respectively. As expected, the estimate of the c-parameter was significantly reduced in the high-load condition, while estimates of the u-parameter were equal in size across load conditions. Hence, the process represented by the u-parameter is characterized by a core feature of automaticity: it is unaffected by variations in cognitive capacity (Bargh 1994). This finding supports the notion that the u-parameter does not merely reflect a deliberate application of US valence irrespective of task instructions. In this study, the standard and reversal conditions were manipulated between participants, precluding a regression analysis of the direct evaluations on the MPT parameter estimates (as in study 1a, only aggregated c- and u-parameter estimates are obtained). Nevertheless, it is illustrative to see how the pattern of EC effects observed on the direct evaluations is consistent with the obtained MPT parameter values on another measure. Remember that in the standard condition, controllable and uncontrollable processes act in concert to produce a regular EC effect. In the reversal condition, both processes act in opposition. While the u-parameter does not differ between the low-load and high-load conditions, the c-parameter is significantly reduced under high load. In the low-load condition, we thus observe a regular EC effect in the standard condition produced by both uncontrollable and controllable processes. As the c-parameter is larger than the u-parameter, controllable processes successfully counteract the influence of uncontrollable ones in the reversal condition, producing a significant, descriptively smaller, reversed EC effect. In the high-load condition, the c-parameter is significantly reduced and thus adds less to the EC effect in the standard condition and less powerfully contradicts uncontrollable processes in the reversal condition. Hence, the regular EC effect in the standard condition is considerably smaller and the opposition of processes in the reversal condition, which are about equally strong under high load, results in a null effect. STUDY 3: MANIPULATION OF MOTIVATION TO CONTROL Study 3 was designed to provide additional evidence for the construct validity of the parameters by manipulating a second crucial variable known to distinguish between controllable and uncontrollable processes—namely, motivation. Consumers’ level of motivation to execute control should affect only controllable processes, leaving uncontrollable processes unaffected (Bargh 1994; Fazio 1990). To investigate the influence of motivation on the controllability of EC, we presented the conditioning phase as a performance task, and manipulated the presence of financial incentives for forming “correct” impressions. Financial incentives should maximize the likelihood that participants try their best to execute the reversals. If truly reflective of a controllable process, the c-parameter should be greater in the financial incentives condition as compared to a control condition. Moreover, in this setup uncontrollable attitude acquisition has negative, undesirable consequences for participants’ financial compensation. To the extent that a significant u-parameter is obtained even in this condition, its interpretation as reflecting uncontrollable evaluative learning will be strengthened. This experiment also serves the secondary goal of strengthening the predictive validity evidence that was obtained in study 1b. In that study, we tested whether the individual c- and u-parameter estimates predict participants’ responses on a separate, continuous measure of evaluation. Whereas the results confirmed our expectations, the sample size in study 1b was relatively small. Therefore, again we implemented a within-participant design and increased the number of participants to estimate the individual parameters with greater accuracy. Method Participants Seventy-four students of different majors at a German university took part in the experiment. Five participants had not understood the instructions (i.e., they executed no reversals in the reversal condition) and were excluded from analysis. This left 69 participants (41 females, 28 males) for analysis (Mage = 24.22 years, SDage = 6.98). Participants received course credit or financial compensation of €3.00. In the financial incentives condition, participants received an additional performance-based reward (see below for details). Participants in the control condition were also paid €2.70 extra (corresponding to 75% correct classifications), but were informed about this only after the experiment. Design The experiment employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: preratings vs. postratings) × 2 (instruction condition: standard vs. reversal) × 2 (financial incentives: control vs. incentives) mixed design with repeated measures on the first three factors. Materials and Procedure The implementation of a performance task required instructions that were more explicit about the usually observed effects of the pairings and participants’ task. We thus implemented instructions that were similar to the standard and reversal instructions used in study 2 (see appendix A for the exact wording of the instructions). Additionally, participants in the financial incentives condition received instructions on a performance-based reward. The standard block was always presented first and followed by the reversal block. We decided on this fixed sequence to reduce error variance and because the sequence factor did not influence the results in study 1b. Results Evaluative Ratings Evaluative ratings, displayed in figure 9, were submitted to a 2 (US valence) × 2 (time of evaluative rating) × 2 (instruction condition) mixed ANOVA. The overall EC effect was significant (F(1, 67) = 5.31, p = .02, ηp2 = .07), as was the main effect of US valence (F(1, 67) = 4.23, p = .04, ηp2 = .06). There also was a significant main effect of time of measurement (F(1, 67) = 4.44, p = .04, ηp2 = .04). The EC effect was moderated by instruction condition (F(1, 67) = 43.45, p < .001, ηp2 = .39), as was the main effect of US valence (F(1, 67) = 32.03, p < .001, ηp2 = .32). None of the other effects was significant (smallest p = .16). FIGURE 9 View largeDownload slide EVALUATIVE RATINGS IN STUDY 3 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND FINANCIAL INCENTIVES CONDITION NOTE.— The error bars represent standard errors. FIGURE 9 View largeDownload slide EVALUATIVE RATINGS IN STUDY 3 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND FINANCIAL INCENTIVES CONDITION NOTE.— The error bars represent standard errors. Separate analyses of the EC effects in the standard and reversal conditions demonstrate that the regular EC effect in the standard condition was significant (F(1, 67) = 52.91, p < .001, ηp2 = .44) and not moderated by incentives condition (F(1, 67) = 1.13, p = .29, ηp2 = .02). While the reversed EC effect in the reversal condition was much smaller than the standard effect, it was also significant (F(1, 67) = 13.69, p < .01, ηp2 = .17), irrespective of incentives condition (F(1, 67) = 1.05, p = .31, ηp2 = .02). As reflected by the partial η², the size of the EC effect in the standard condition was more than twice as large as the size of the (reversed) EC effect in the reversal condition. This size difference was statistically significant (F(1, 68) = 5.45, p = .02, ηp2 = .07). MPT Model The initial model containing one c-, one u-, and one r-parameter per incentives condition demonstrated a good fit to the data (G²(2) = 0.55, p = .76) and is presented in figure 10. The initial parameter estimates in the control condition were c = .33, 95% CI [.26, .39], u = .10, 95% CI [.00, .19] (significantly different from zero, ΔG²(1) = 4.24, p = .02), and r = .49, 95% CI [.44, .55]. The parameter estimates for the financial incentives condition were c = .61, 95% CI [.55, .66], u = .15, 95% CI [.00, .29] (significantly different from zero, ΔG²(1) = 4.03, p = .02), and r = .50, 95% CI [.42, .59]. While the u-parameters did not differ between incentives conditions (ΔG²(1) = 0.33, p = .57), the c-parameters were different (ΔG²(1) = 42.61, p < .001). The r-parameters also did not differ (ΔG²(1) = 0.03, p = .86). FIGURE 10 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 3 NOTE.— c denotes the controllable parameter with separate estimates for the control and incentives conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars indicate the standard error of the individual parameter estimates. FIGURE 10 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 3 NOTE.— c denotes the controllable parameter with separate estimates for the control and incentives conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars indicate the standard error of the individual parameter estimates. Predictive Value of Individual Parameter Estimates As in study 1b, we conducted an HLM predicting the EC effect as a function of participants’ individual c- and u-parameter estimates (standardized), the instruction condition, the financial incentives condition, and the parameters’ interactions with the instruction and incentives conditions. For 10 participants, the u-parameter was unidentified (this occurs when some category frequencies are too close to zero). The main effect of instruction condition was again significant (B = –35.94, SE = 4.92, t(54) = –7.31, p < .001), reflecting the fact that the EC effect is positive in the standard condition, but negative in the reversal condition. The main effect of the c-parameter was marginally significant (B = –4.60, SE = 2.37, t(53) = –1.94, p = .058), but this effect was moderated by instructions (B = –23.57, SE = 5.50, t(54) = –4.28, p < .001). As expected, the c-parameter showed a positive relation to the EC effect in the standard condition (B = 7.19, SE = 3.63, t(53) = 1.98, p = .05) and a negative relation in the reversal condition (B = –16.38, SE = 3.63, t(53) = –4.51, p < .001). The u-parameter was a universally positive predictor of the EC effect (B = 8.64, SE = 2.04, t(53) = 4.24, p < .001), as its effect was not moderated by instruction condition (B = 1.54, SE = 4.80, t(54) = 0.32, p = .75). Finally, the financial incentives manipulation did not have a direct effect on EC, nor did it qualify the (main or interactive) effects of the c- and u-parameters on EC (smallest p = .13). Discussion This study provides several important pieces of evidence for the existence of uncontrollable EC effects. First, the study strengthens the predictive validity of the MPT parameters. As in study 1b, the HLM analysis indicates that the c- and u- parameters correlate as predicted with an independent measure of the EC effect. Second, the study assuages a potential concern that participants in previous studies might not have been sufficiently motivated to comply with task instructions. In this study’s financial incentives condition, it was financially costly for participants to acquire evaluations in an uncontrollable manner. Yet, even in this condition, EC continued to exert an uncontrollable influence on the CS evaluations, as evidenced by the significant u-parameter. Third, and most importantly, the study enhances our faith in the construct validity of the u- parameter as capturing an uncontrollable process. As demonstrated by its effect on the c-parameter, the financial incentives manipulation did increase participants’ likelihood to exert control. At the same time, however, financial incentives did not influence the u-parameter. This implies that in those cases where control fails, there remains a constant probability of an uncontrollable effect of EC, unaffected by the motivation to exert control. This process’s resistance to effortful control supports the notion that this parameter does not represent the deliberate application of US valence regardless of task instructions and is a hallmark property of its automatic nature. In this study, we were highly explicit about the performance criterion necessary to achieve additional rewards. While this study supports the existence of an uncontrollable learning mechanism, the study’s emphasis on performance imposes a limitation on the interpretation of the c-parameter. It is possible that in some cases, the c-parameter does not capture genuine evaluative learning, but rather a demand effect of participants reporting the correct CS valence as requested by the instructions. Demand artifacts are a notable source of concern in many experimental settings, and particularly so in settings where the researcher’s hypotheses and expectations are transparent for participants (Page 1969). Note, however, that one desirable and crucial advantage of the method we developed is that while demand effects could inflate the size of the c-parameter, they go against finding a significant u-parameter, which is derived from participants’ inability to conform to instructions or task demands. STUDY 4: EXTERNAL VALIDITY OF PARADIGM The purpose of the next study was to increase the external validity of our paradigm in two different ways. All previous studies used the same stimulus materials, working with naturalistic faces as CSs. In the next study, we employed consumer brands as CSs to investigate whether our results hold in the consumer context, as we would claim. Second, our previous studies were primarily concerned with the internal validity of our research paradigm. To demonstrate the internal validity of this research, it is important that participants are instructed to exert control to the best of their ability. This is the reason that in all previous studies, we explicitly encouraged participants to either apply or reverse the valence of the US in forming their attitude toward the CS. As is often the case, however, maximizing internal validity comes at the cost of reducing the external validity—that is, the generalizability of the findings to other paradigms and settings. Specifically, the instructions in EC research more generally often conceal the purpose of the CS-US pairings (Olson and Fazio 2001; Sweldens et al. 2010). The question then arises whether our method is also informative about such contexts, and what would be the estimated contribution of controllable and uncontrollable processes when participants are not instructed to exert control. Analogous to the previous studies, we included a standard and reversal condition with highly explicit instructions to apply or reverse the influence of US valence, respectively. Additionally, we included a “covert condition” with instructions that conceal the purpose of the pairings, thereby reducing the probability that participants exert control over the attitude acquisition. Method Participants We collected data of 173 international students at Erasmus University Rotterdam. We excluded two participants (1.16 %) who gave the same response in the MPT task in 88% of the cases and above. Of these participants, 94 were women and 97 were men. They were between 17 and 24 years old (Mage = 19.94, SDage = 1.25). Design The experiment implemented a 2 (US valence: positive vs. negative) × 2 (time of measurement: pre- vs. post-pairings) × 3 (instructions: standard vs. reversal vs. covert) with repeated measures on the first two factors. Materials and Procedure The CS repertoire consisted of 38 bottled water brand logos that were specifically designed for this experiment and thus unknown to participants. For each participant, the software selected the 24 CSs that received the most neutral preratings. The instructions were adapted to the materials and were very explicit about the goal of the task (exerting control) in the standard and reversal conditions. In the covert condition, we employed instructions that were used by Sweldens and colleagues (2010). These instructions asked participants to familiarize themselves with the brands and concealed the purpose of the pairings by stating that the US images were presented to make the presentation of the CS brands more interesting (the exact wording can be found in appendix A). Results Evaluative Ratings The evaluative ratings revealed a large EC effect (F(1, 168) = 65.17, p < .001, ηp2 = .28) that was moderated by condition (F(2, 168) = 7.93, p < .001, ηp2 = .09) as apparent in figure 11. The EC effect was large in both the covert (F(1, 56) = 36.08, p < .001, ηp2 = .39) and the standard conditions (F(1, 54) = 31.31, p < .001, ηp2 = .37) and was reduced to a small but regular EC effect in the reversal condition (F(1, 58) = 4.14, p < .05, ηp2 = .07). The absolute sizes of the EC effects did not differ between the standard and covert conditions (F(1, 110) = 2.05, p = .16, ηp2 = .02), while effect sizes differed between the standard and reversal conditions (F(1, 112) = 13.57, p < .001, ηp2 = .11), as well as the covert and the reversal conditions (F(1, 114) = 8.24, p < .01, ηp2 = .07). FIGURE 11 View largeDownload slide EVALUATIVE RATINGS IN STUDY 4 BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. FIGURE 11 View largeDownload slide EVALUATIVE RATINGS IN STUDY 4 BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. The pattern of means across the three conditions suggests that in this experiment, the instructions mostly affected the positively paired CSs. Indeed, the moderation of the EC effect by instruction condition was significant only for positively paired CSs (F(2, 168) = 8.20, p < .001, ηp2 = .09), but not for negatively paired CSs (F(2, 168) = 1.99, p = .14, ηp2 = .02). This suggests that control was not successfully applied when US images were negative. The MPT model will test for this assumption by comparing the c-parameters between the positive and negative US valence conditions. MPT Model We first analyzed the MPT model for the standard and reversal conditions only. The estimates of the c-parameter differed significantly between positive and negative US valences, ΔG²(1) = 16.05, p < .001. These parameters thus must be estimated separately. The c-parameter in the negative pairings condition amounted to c- = .05, 95% CI [.00, .10], and in the positive pairings condition to c+ = .20, 95% CI [.15, .25]. The estimate of the u-parameter was u = .18, 95% CI [.14, .22], and differed significantly from zero, ΔG²(1) = 70.02, p < .001. The r-parameter amounted to r = .53, 95% CI [.51, .56]. This slight bias toward “pleasant” was significant, ΔG²(1) = 5.84, p = .02. With four parameters this model is saturated and thus does not allow for an assessment of fit, but the fact that the G² statistic was not larger than zero indicates that this model’s equality restrictions (e.g., that we expected more responses to oppose US valence in the reversal condition) were met, G²(0) = 0.00. The parameter estimates are depicted in figure 12. FIGURE 12 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 4 NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 12 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 4 NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. In the next step, we calculated the MPT model for all three experimental conditions, including the covert condition with the more typical EC instructions. The u-parameter amounted to u = .16 [.12, .20], and was significantly larger than zero, ΔG²(1) = 68.39, p < .001. Furthermore, it did not differ significantly from its estimate when we used the data of the standard and reversal conditions only, ΔG²(1) = 1.50, p = .22. The estimate of the u-parameter was thus not altered by the inclusion of the 1,368 observations from the covert condition. Again we estimated the c-parameters of the standard and reversal conditions separately for negative, c- = .04, 95% CI [.00, .10], ΔG²(1) = 2.70, p = .05, and positive pairings, c+ = .19, 95% CI [.14, .24], ΔG²(1) = 54.03, p < .001, as they differed strongly from another, ΔG²(1) = 15.98, p < .001. The estimate of the c-parameter in the covert condition amounted to ccovert = .00, 95% CI [.00, .08], and did not differ from zero, ΔG²(1) = 0.00, p = 1. The r-parameter amounted to r = .53, 95% CI [.51, .55], and thereby indicated a slight tendency to respond “pleasant” in this task, ΔG²(1) = 9.46, p < .01. With the addition of the covert condition, we gained degrees of freedom allowing for an assessment of model fit. The final model across all conditions described the data well, G²(1) = 2.66, p = .10. Discussion The present study allows several comparisons between explicit control instructions and instructions that covered the true purpose of the pairings. When calculating a joint model for all three conditions, we restricted the u-parameter to be equal across conditions. Nevertheless, its value could have changed due to the inclusion of 1,368 additional observations. The fact that its estimate was not affected is consistent with its interpretation as reflecting the uncontrollable influence of US pairings. Moreover, the c-parameter that could vary freely was strongly affected by the presence versus absence of explicit control instructions, as it was reduced to insignificance in the covert condition, where participants were not instructed to exert control. Hence, the present results suggest that uncontrollable learning may be relatively stable across different EC paradigms, while the degree of controllable learning may vary considerably. The fact that the MPT parameters respond as predicted to the presence versus absence of control instructions is further evidence for their construct validity. Furthermore, the study provides evidence for the relevance of our approach to the wider EC literature. The EC effect in the standard and covert conditions was approximately equal in size, assuaging concerns that the EC effect in our studies might be completely dependent on the presence of explicit control instructions. In fact, our data even indicate that controllable processing might play a negligible role in typical EC paradigms, as evidenced by the fact that ccovert did not differ from zero. One finding we did not anticipate was the difference in control exertion between positive and negative US pairings, which was not present in any other study. The fact that participants exerted less control over negative US pairings might be consistent with the so-called “negativity bias,” in which negative information has a stronger effect on evaluations than comparable positive information (Ito et al. 1998). We do not know, however, why this bias manifested itself solely in this study and could only speculate about the various slight differences in sample characteristics and in the combination of conditioned stimuli (consumer brands) and instructions unique to this study (see appendix A for the differences). STUDY 5: PREDICTING PRODUCT CHOICE AND CONSUMPTION This study’s main goal was to investigate whether the uncontrollable EC effect on attitudes documented in the previous studies influences actual product choices and consumption. Apart from its primary substantive importance, finding evidence for a relation between the MPT model parameters and actual product choices would establish a crucial part of the predictive validity of the MPT modeling approach. To study the effects on product choices, we again used the initially neutral logos of bottled water brands as our CS repertory, which would be paired with positive versus negative pictures during the conditioning phase. We then assessed the choices of participants when they were given the opportunity to taste these waters. Moreover, we also assessed the amount they consumed as an unobtrusive measure of approach behavior. Method Participants Eighty-one students of different majors at the University of Tübingen took part in this experiment (53 female; Mage = 23.14, SDage = 4.87). The MPT modeling was based on 1,944 observations. Design The experiment employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: preratings vs. postratings) × 2 (instruction condition: standard vs. reversal) design with repeated measures on all three factors. Sequence of instructions (standard-first vs. reversal-first) was counterbalanced between participants. Materials and Procedure With the exception that the covert condition was omitted from this study, materials and procedure were identical to the previous study. Going beyond the previous study, participants were invited to refresh themselves with the waters of the brands presented during the study in an adjacent room after the computer-based part was finished. Plastic cups that were labeled with the CS logos were lined up on a table. Each cup was filled with 130 milliliters of tap water. Participants were told that they could drink as little or as much from the cups as they liked. An experimenter who was blind to the assignment of CSs to conditions recorded the cups selected by each participant. After the participant left, the amount consumed was determined with a calibrated digital weighing scale. Results Evaluative Ratings The EC effect was significant, F(1, 80) = 28.77, p < .001, ηp2 = .27, but qualified by instruction condition, F(1, 80) = 88.78, p < .001, ηp2 = .53. Both the regular EC effect in the standard condition, F(1, 80) = 119.54, p < .001, ηp2 = .60, and the reversed EC effect in the reversal condition, F(1, 80) = 6.87, p <.05, ηp2 = .08, were significant. The difference in the absolute sizes of the EC effect was significant, F(1, 80) = 28.77, p < .001, ηp2 = .27 (figure 13). FIGURE 13 View largeDownload slide EVALUATIVE RATINGS IN STUDY 5 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND FINANCIAL INCENTIVES CONDITION. NOTE.— The error bars represent standard errors. FIGURE 13 View largeDownload slide EVALUATIVE RATINGS IN STUDY 5 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND FINANCIAL INCENTIVES CONDITION. NOTE.— The error bars represent standard errors. Consumption Behavior We analyzed the number of cups chosen and the total amount of water consumed in two separate 2 (US valence) × 2 (instruction condition) repeated-measures ANOVAs as displayed in figure 14. Participants’ choices of cups in the tasting procedure demonstrated a main effect of US valence (F(1, 80) = 24.14, p < .001, ηp2 = .23), a main effect of instruction condition (F(1, 80) = 5.40, p = .02, ηp2 = .06), and their interaction (F(1, 80) = 72.89, p < .001, ηp2 = .48). Planned contrasts revealed that participants selected more cups of positively than negatively paired brands in the standard condition (F(1, 80) = 94.26, p < .001, ηp2 = .54), which was reversed in the reversal condition (F(1, 80) = 11.68, p < .001, ηp2 = .13). Results from the consumption measure were equivalent. The total amount of water consumed was a function of US valence × instruction condition (F(1, 80) = 53.06, p < .001, ηp2 = .40), indicating a regular EC effect in the standard condition (F(1, 80) = 61.41, p < .001, ηp2 = .43) and a reversed EC effect in the reversal condition (F(1, 80) = 7.74, p < .01, ηp2 = .09). FIGURE 14 View largeDownload slide BEHAVIORAL MEASURES TAKEN IN STUDY 5 BY US VALENCE AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. FIGURE 14 View largeDownload slide BEHAVIORAL MEASURES TAKEN IN STUDY 5 BY US VALENCE AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. MPT Model The parameter estimates and standard errors are displayed in figure 15. They amounted to c = .48, 95% CI [.44, .51], u = .38, 95% CI [.31, .45], and r = .47, 95% CI [.41, .52]. The u-parameter was larger than zero, ΔG²(1) = 101.35, p < .001. The model fit the data well (G²(1) = 1.62, p = .20; frequency data in appendix B). FIGURE 15 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 5 NOTE.— c denotes the controllable parameter with separate estimates for the high-load and the low-load conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 15 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 5 NOTE.— c denotes the controllable parameter with separate estimates for the high-load and the low-load conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. Predictive Value of Individual Parameter Estimates We conducted three HLMs to assess the (standardized) c- and u-parameters’ impact on the attitudinal EC effect (i.e., the difference in attitudes between positively and negatively conditioned brands), the brand choice EC effect (i.e., the difference in cups chosen of these brands), and the consumption EC effect (i.e., the difference in water consumption between these brands). Six participants had unidentified u- and r-parameters (this occurs when some category frequencies are close to zero). The analysis on evaluative ratings revealed a significant main effect of instruction condition, as the EC effect is positive in the standard condition, but reversed in the reversal condition (B = –64.46, SE = 5.25, t(71) = –12.29, p < .001). We also observed a significant main effect of the c-parameter (B = –6.93, SE = 2.73, t(71) = –2.54, p = .01), which was moderated by instruction condition (B = –39.29, SE = 5.57, t(71) = –7.06, p < .001). While the c-parameter predicted a regular EC effect in the standard condition (B = 12.71, SE = 3.90, t(71) = 3.26, p < .01), it generated a reversed EC effect in the reversal condition (B = –26.58, SE = 3.90, t(71) = –6.82, p < .001). Conversely, the u-parameter had a universally positive influence on the EC effect (B = 17.72, SE = 2.59, t(71) = 6.84, p < .001), which was not moderated by reversal instructions (B = –3.14, SE = 5.29, t(71) = –0.59, p = .55). The analysis on brand choice revealed a main effect of instruction condition (B = –2.80, SE = 0.29, t(71) = –9.56, p < .001) and a main effect of the c-parameter (B = –0.31, SE = 0.12, t(71) = –2.59, p = .01). Importantly, the latter was qualified by instruction condition (B = –1.60, SE = 0.31, t(71) = –5.13, p < .001). As expected, the c-parameter had a positive influence on brand choice in the standard condition (B = 0.49, SE = 0.20, t(71) = 2.48, p = .02) and a negative influence in the reversal condition (B = –1.11, SE = 0.20, t(71) = –5.64, p < .001). Again, the u-parameter had a universally positive influence (B = 0.62, SE = 0.11, t(71) = 5.41, p < .001), unaffected by instructions (B = –0.09, SE = 0.30, t(71) = –0.31, p = .76). The analysis of water consumption (in milliliters) mirrored the analysis on brand choice. Again, the u-parameter had a universally positive influence (B = 12.55, SE = 5.62, t(71) = 2.23, p = .03), unaffected by instructions (B = 3.21, SE = 10.58, t(71) = 0.30, p = .76). The c-parameter’s effect was moderated by instructions (B = –43.54, SE = 11.13, t(71) = –3.91, p < .001), with a positive weight in the standard condition (B = 13.35, SE = 8.12, t(71) = 1.64, p = .10) and a negative weight in the reversal condition (B = –30.19, SE = 8.12, t(71) = –3.72, p < .001). Discussion Study 5 goes beyond the demonstration of uncontrollable evaluative learning and establishes the predictive validity of the MPT parameters for product choice and consumption. The present results clearly demonstrate the uncontrollable impact of visual affective stimulus pairings on consumption behavior that comprised both the choice of drinks and their consumption. Participants often selected and consumed brands that were presented with positive contents, and did not consume brands that were presented with negative images, even though they were instructed to reverse the impact of the pairings on their evaluations. Hence, even when participants knew about the deceptive nature of advertising, neither their evaluations nor their consumption behavior were immune to its content. SAMPLE SIZES AND POWER IN STUDIES 1–5 Sample sizes were based on previous research using MPT modeling (Hütter and Sweldens 2013; Hütter et al. 2012). In all studies, each participant provided 24 observations (i.e., responses towards 24 CSs) as a basis for estimating the MPT model. In the first experiments testing merely the structure of the model, we thus aimed for up to 40 (study 1a) and 20 participants (study 1b), respectively. These sampling strategies resulted in 480 observations even in the study with the smallest sample size (study 1b; see also appendix B). We increased sample size considerably in all studies that introduced between-participants factors and those that assessed correlations between measures. As apparent from table 1, power analyses support our substantive conclusions in all experiments. Even in study 1b, the power to detect a u-parameter amounted to .70 and the power of detecting an influence of uncontrollable learning exceeds the conventional standard of .80 in all the other studies. Furthermore, the lack of an effect of attentional load (study 2) and financial incentives (study 3) on the u-parameter cannot be explained by a weak or unsuccessful manipulation, as they clearly affected the c-parameter (and evaluative ratings in study 2). The effect sizes of the manipulations’ effects on the c-parameters were in all cases considerably larger than the effects on the u-parameters, which were negligible. Table 1 Power Analyses for Central Hypothesis Tests Significance of u-parameter (test against zero) Study Parameter Cohen’s w N Power (1 – β) 1a u 0.11 888 .90 1b u 0.11 480 .70 2 u 0.07 1656 .80 3 u 0.08 1416 .85 4 u 0.17 4104 1 5 u 0.23 1644 1 Parameter comparisons across designs and experimental conditions 1a/1b: within versus between c 0.06 1368 .56 u 0.01 .06 2: low versus high load c 0.16 1656 1 u 0.01 .09 3: control versus incentives c 0.11 1416 .98 u 0.02 .15 Significance of u-parameter (test against zero) Study Parameter Cohen’s w N Power (1 – β) 1a u 0.11 888 .90 1b u 0.11 480 .70 2 u 0.07 1656 .80 3 u 0.08 1416 .85 4 u 0.17 4104 1 5 u 0.23 1644 1 Parameter comparisons across designs and experimental conditions 1a/1b: within versus between c 0.06 1368 .56 u 0.01 .06 2: low versus high load c 0.16 1656 1 u 0.01 .09 3: control versus incentives c 0.11 1416 .98 u 0.02 .15 NOTE.— N = number of observations. The rule of thumb for effect size w is that w = .1 is “small,” w = .3 is “medium,” and w = .5 is “large” (Cohen 1988). Power analyses were conducted with multiTree (Moshagen 2010). Table 1 Power Analyses for Central Hypothesis Tests Significance of u-parameter (test against zero) Study Parameter Cohen’s w N Power (1 – β) 1a u 0.11 888 .90 1b u 0.11 480 .70 2 u 0.07 1656 .80 3 u 0.08 1416 .85 4 u 0.17 4104 1 5 u 0.23 1644 1 Parameter comparisons across designs and experimental conditions 1a/1b: within versus between c 0.06 1368 .56 u 0.01 .06 2: low versus high load c 0.16 1656 1 u 0.01 .09 3: control versus incentives c 0.11 1416 .98 u 0.02 .15 Significance of u-parameter (test against zero) Study Parameter Cohen’s w N Power (1 – β) 1a u 0.11 888 .90 1b u 0.11 480 .70 2 u 0.07 1656 .80 3 u 0.08 1416 .85 4 u 0.17 4104 1 5 u 0.23 1644 1 Parameter comparisons across designs and experimental conditions 1a/1b: within versus between c 0.06 1368 .56 u 0.01 .06 2: low versus high load c 0.16 1656 1 u 0.01 .09 3: control versus incentives c 0.11 1416 .98 u 0.02 .15 NOTE.— N = number of observations. The rule of thumb for effect size w is that w = .1 is “small,” w = .3 is “medium,” and w = .5 is “large” (Cohen 1988). Power analyses were conducted with multiTree (Moshagen 2010). GENERAL DISCUSSION Consumers are constantly exposed to advertisements featuring visual imagery with affective content, which comes as no surprise, as it has long been established that pairing brands with appetitive stimuli is an effective method to increase brand liking (Bierley et al. 1985; Stuart et al. 1987). Very little research exists that investigates the controllability of attitude acquisition via such pairings. In the current research, we develop and validate a method to separate and quantify controllable and uncontrollable attitude acquisition in evaluative conditioning. We report six experiments, which demonstrate that simple presentations of visual affective stimuli can have uncontrollable effects on consumers’ attitudes and behavior. We discuss the theoretical, methodological, and substantive implications of our findings, as well as their limitations. Methodological Contributions Over the past two decades, consumer research has had a strong focus on—and made great progress in—studying implicit processes in memory, affect, and persuasion (Johar, Maheswaran, and Peracchio 2006). Standards for the investigation of implicit processes have been discussed intensely recently (Sweldens et al. 2017; Williams and Poehlman 2017). Even though the processing tree framework has many advantages over conventional approaches, applications thereof have been surprisingly uncommon in consumer research, with a few notable exceptions (Dalton and Huang 2014; Fitzsimons and Williams 2000; Pham and Johar 1997; Shapiro 1999). Nevertheless, further progress can still be made and a more widespread adoption of the processing tree framework, accompanied with more up-to-date methods of analysis, has the potential to improve our knowledge in this fascinating subject area significantly. Challenges of Previous Approaches to Uncontrollability Only very recently, research has started to explore whether pairings with visual affective stimuli can have uncontrollable effects on attitudes. Two articles by Gawronski and colleagues (2014, 2015) demonstrated that EC procedures have uncontrollable effects on indirect attitude measures, while effects on direct measures could still be controlled. As we outlined in more detail in the introduction, we believe more research on this topic is necessary for both substantive and methodological reasons. Substantively, EC effects are fully uncontrollable when they fail to be corrected at both the encoding stage and the validation stage—that is, at retrieval and application of these evaluations as conceived in two-step frameworks of attitudes (Fazio 1990; Gawronski and Bodenhausen 2006). We thus focused on direct measures of evaluation that allow participants to exert full control over the retrieval and application of their evaluations. Theoretically and methodologically, it can be treacherous to rely on reaction-time-based indirect measures, as these measures are insensitive to negations (Deutsch et al. 2006). Hence, finding unmodulated EC effects on indirect measures could be a spurious indicator of uncontrollability. The methodological conundrum at the heart of this debate is that neither indirect nor direct measures of evaluation are process-pure reflections of automatic or nonautomatic processes. Each will reflect the contribution of both types of processes to varying degrees. Therefore, it is possible that the apparently controllable nature of direct evaluations observed in previous research masked an uncontrollable effect on some evaluations. To distinguish automatic from nonautomatic contributions to a measurement outcome, one needs to apply processing tree modeling—the primary undertaking of the current research. The Processing Tree Approach We developed an MPT model to distinguish controllable from uncontrollable EC effects in a direct evaluative measure, estimated by the model’s c- and u-parameters, respectively. Its main purpose was to investigate the potential existence of an uncontrollable EC effect on direct evaluations. A general challenge with developing a new measurement method is to demonstrate that it produces valid indicators of the targeted constructs. We have tested the validity of the method through a triangulation of approaches in our studies. Specifically, our experiments provided construct validity for our conceptualization of the c- and u-parameters as capturing an effortful, controllable process versus an efficient, uncontrollable process, respectively. Furthermore, across different EC procedures we demonstrated the external validity of our findings. We provided predictive validity by demonstrating that the parameter estimates were predictive of EC effects as assessed on independent evaluative measures as well as consumers’ choice and consumption behavior. Does the u-Parameter Reflect Uncontrolled or Uncontrollable Applications of US Valence? Readers may wonder to what degree our findings might be more reflective of the participants’ failure to exert control successfully rather than of the role of uncontrollable learning. This question may be approached from two perspectives, a conceptual and a methodological one. Conceptually, our theorizing about the role of controllable and uncontrollable processes as well as unsystematic processes is formalized in the MPT model. In that model, the failure of control (1 – c) and uncontrollable learning (u) are two orthogonal events. That is, if individuals fail to comply with task instructions with the probability 1 – c, there are two mutually exclusive outcomes. The first possible outcome is that participants applied US valence to the CS irrespective of task instructions (u). The second possible outcome is that participants did not learn uncontrollably (1 – u). In that case, they would (randomly) indicate that the CS is either positive (with the probability r) or negative (with the probability 1 – r). In other words, not following the task instructions or being incapable of doing so does not automatically lead to EC effects captured by the u-parameter. In the absence of controllable learning (1 – c), it is still important to investigate whether there is also uncontrollable learning (u). Only if a systematic process operates by which US valence is transferred to the CSs would the overall u-parameter be larger than zero. Moreover, US valence may be applied to the CSs irrespective of task instructions in an uncontrolled rather than an uncontrollable manner. For instance, participants may have lacked the capacity or motivation to exert control and may have applied US valence in a deliberate manner regardless of instructions. Therefore, it was important to investigate the u-parameter’s sensitivity to cognitive load (study 2), financial incentives (study 3), and control instructions (study 4). These experiments demonstrate the insensitivity of the u-parameter to all of these factors, demonstrating that it does not predominantly reflect deliberate processes. The present research thus attests to the partly uncontrollable nature of the EC effect. What Are the Effects of Potential Violations of the Invariance Assumption? The estimation of the MPT parameters rests on the assumption that the contributions of the latent processes are invariant across instruction conditions (Jacoby 1991). Previous research has shown that this assumption is often not met for the dominant process and that its violation biases the parameter estimation of the nondominant process (Klauer et al. 2015). In the present research, the invariance assumption for the dominant process implies that the contribution of the controllable process should be constant across standard and reversal conditions. If this assumption were violated, the estimation of the uncontrollable process would be distorted. As our designs preclude a direct, empirical test of this assumption, we cannot dismiss the possibility that the u-parameter was influenced by its violation. Yet several pieces of evidence speak against the artifactual nature of this parameter. Let us discuss one by one the two possible ways in which the contributions of the controllable process could differ between standard and reversal conditions. On the one hand, controllable processing may be instigated only by difficult tasks, in which case participants would have exerted more control in the reversal condition than in the standard condition. This notion is in line with demonstrations of increasing degrees of analytical reasoning with increasing perceptions of difficulty (Alter et al. 2007), with the conflict monitoring hypothesis that states that the degree of cognitive control is tuned to the degree of conflict (Botvinick et al. 2001), and with our observation that no control was exerted in the covert condition of study 4. If more control were exerted under reversal instructions, the contribution of the controllable processes would be overestimated, because the model would match the successfully reversed EC cases in the reversal condition with cases of CSs that acquired a regular EC effect in the standard condition. Thus, the model would treat uncontrollably acquired EC effects in the standard condition as if they were due to controllable learning. In this case, the role of uncontrollable learning would be even larger than documented in the present article. The most threatening case for the current research would arise when controllable processing plays a smaller role under reversal instructions than in the standard condition, because it could artificially inflate the u-parameter. Let us illustrate this case by means of a simplification of the modeling logic and under the assumption that there is no uncontrollable learning. If more control were exerted in the standard condition than in the reversal condition, the model would underestimate the amount of controllable learning in the standard condition, as it assumes that its contribution is as large as in the reversal condition. As this underestimation would leave unexplained cases in the standard condition that are in line with US valence, they would feed the u-parameter even though they are the result of controllable learning. Moreover, because an invariance assumption also applies to the uncontrollable component, the model would assign cases that are due only to unsystematic processes (“response tendencies” or “guessing”) in the reversal condition to the u-parameter. In this case, the u-parameter would only be an artifact reflecting a mixture of controllable processes in the standard condition and unsystematic processes in the reversal condition and have no bearing on uncontrollable evaluative learning. Several empirical findings alleviate such a concern. First, the c- and u-parameters demonstrate discriminant validity; that is, they react differently to the experimental manipulations as predicted on conceptual grounds. If the u-parameter’s estimate were dependent on the difference of control exerted in the standard versus reversal conditions, these clear-cut patterns would have been very unlikely. A second counterargument to this alternative explanation stems from the HLMs regressing the EC effect on the c- and u-parameter estimates in the within-participant designs. If the u-parameter comprised unsystematic processes in the reversal condition, it should not meaningfully predict the effects of the positive and negative stimulus pairings on independently assessed measures of the EC effect beyond the (opposite) predictive value of the c-parameter. Supporting the notion of a genuine learning process, the u-parameter predicts direction and size of the EC effect across standard and reversal conditions in a consistent manner. Moreover, its predictive value is further established for behavioral data. In summary, while we cannot exclude the possibility of violations of the invariance assumption influencing the estimate of the u-parameter, at the same time our findings speak against the notion that the u-parameter is merely an artifact of assumption violations. Theoretical Contributions Whether EC can have uncontrollable effects is part of a wider debate regarding the potential existence of automatically operating processes in associative learning. While the existence of implicitly operating processes in memory retrieval is generally uncontested, this is not the case for learning or encoding of associations (Mitchell, De Houwer, and Lovibond 2009; Shanks 2010). Despite decades of research, mostly focusing on the necessity for “contingency awareness” (another feature of automaticity), closure on this matter has not been reached (Sweldens et al. 2014). Studying the uncontrollability of evaluative learning has the potential to provide a fresh perspective on this long-standing debate. This presumes, however, that the MPT model is effective at separating nonautomatic from automatic processes. For the u-parameter in particular, this poses additional challenges. Technically, the u-parameter provided by the MPT model is an estimate of the percentage of trials where CS evaluations are in line with US valence—irrespective of the participant’s intentions or instructions—after accounting for trials that can be explained by controllable processes (which should be contingent on instructions by definition) and unsystematic processes. In other words, u is an estimate of uncontrolled CS evaluations. To make the inference leap from uncontrolled to uncontrollable, what is left to demonstrate is that this parameter is characterized by a relative independence of processing resources and motivation (Bargh 1994; Fazio 1990). Study 2 was designed to assess the effect of variations in cognitive processing capacity on the MPT parameter estimates. As expected, a reduction in cognitive resources had a detrimental effect on the likelihood to successfully exert control (c), yet the likelihood that US valence is transferred regardless of instructions remained constant and significant (u). Similarly, study 3 was designed to assess the effect of variations in motivation to exert control, by making participants’ payout contingent on their ability to follow instructions. As expected, a surge in motivation increased the likelihood that control can be successfully exerted (c). Again, however, there remained a constant likelihood that US valence would be transferred regardless of instructions (u). These data corroborate the hypothesis that the c- and u-parameters reflect the contribution of controllable and uncontrollable processes, respectively. The finding of a significant u-parameter in all our studies therefore testifies to the partially uncontrollable and automatic nature of EC effects, providing an important contribution to this debate. Substantive Implications In the US alone, companies spend $190 billion annually on advertising, a figure that continues to rise. Participants in our studies went through a series of trials in which initially neutral stimuli were paired with visual affective stimuli, resembling the experience of consumers processing a sequence of advertisements where brands or products are often linked with positive stimuli such as images and music. Our findings imply that a significant percentage of ads will change their evaluations in an uncontrollable fashion, thereby threatening their autonomy. From a consumer protection point of view, it would be vital to assess the magnitude of this effect and understand its prevalence. One powerful feature of the MPT methodology is that it provides quantitative probability estimates of nonautomatic versus automatic processes in a given setting. As revealed by the c-parameter, the ability to exert control over EC effects is far from perfect. When consumers are strongly motivated (e.g., by financial incentives in study 3), their success rate can be increased. At the same time, even a relatively simple secondary task decreased the success rate considerably (study 2). From a consumer protection point of view, this is worrying. As consumers rarely process advertisements with undivided attention, this implies they are generally unlikely to exert control. It is important to realize, however, that a failure to exert control does not automatically imply the presence of an uncontrollable EC effect. When control fails, it remains possible that no valence is transferred and the consumer’s attitude remains unchanged. However, the MPT model indicated a relatively constant likelihood that an uncontrollable EC effect would be established when control failed. It can be concluded that mere co-occurrences of brands with visual affective stimuli will often have uncontrollable effects on consumers’ evaluations and behavior despite their best efforts to correct for this influence. This could have important implications for public policy and regulation of advertising. The concern is especially acute in domains where consumers make decisions that directly affect their health and well-being. One prominent example is advertising for foods or pharmaceuticals. As outlined by Biegler and Vargas (2013), rather than being outlawed as in many other jurisdictions, in the US pharmaceutical advertising is regulated by the Food and Drug Administration, which prohibits the use of “false” or “misleading” claims in pharmaceutical messages. However, as also noted by these authors, regulations currently only verify the propositional content of these messages (textual, verifiable claims), whereas the persuasive power of these messages might rather stem from their unverifiable content (i.e., the images and musical elements), precisely the type of content our research indicates has uncontrollable effects on consumers’ attitudes. Limitations and Future Directions An important limitation of the current findings is that the exact psychological mechanisms underlying the parameters remain unclear. We obtained evidence that the process underlying the u-parameter is characterized by more features of automaticity than the drivers of the c-parameter. However, the u-parameter is still compatible with several theoretical accounts—for example, with automatically operating association formation processes (Baeyens et al. 1992) but also with implicit misattribution or “direct transfer” of evaluative responses (Jones, Fazio, and Olson 2009; Sweldens et al. 2010). One interesting direction would be to manipulate properties of the conditioning procedure, which make certain processes (e.g., implicit misattribution) more likely. Related to the previous limitation, the present research did not investigate the degree to which uncontrollable attitude acquisition shows other features of automaticity besides efficiency and unintentionality, such as unawareness or memory independence (with the latter often being used as a proxy for unawareness). We support a disjunctive view of automaticity, meaning that we do not expect all features to be either present or absent in a given EC procedure (Bargh 1994; Fiedler and Hütter 2014). Specifically, on the one hand, there may be instances where individuals are fully aware of the CS-US pairings yet are unable to control the resulting attitude. In fact, given the explicit instructions, we assume that awareness levels were very high in the present research. On the other hand, uncontrollable attitude acquisition may be facilitated by unawareness of the pairings. Such questions need to be addressed empirically. Future research should also extend our knowledge on the boundary conditions for controllable learning. For example, in the present setting participants are not provided with any information on how CSs and USs relate to each other. Moreover, participants receive no guidance on specific strategies to exert control. It would be interesting to provide consumers with different strategies to exert control and investigate their effectiveness with a sensitive MPT procedure, extending the work by Gawronski and colleagues (2015). Conclusion The method developed in this article has made it possible to test for the existence of an uncontrollable EC effect on explicit evaluations, to quantify the magnitude of its impact, and to demonstrate its downstream consequences on product choice and consumption. Thereby, this research provides an important contribution to our understanding of the acquisition of consumer attitudes. Our experiments provide evidence for independently operating controllable and uncontrollable processes. This might have important practical implications for safeguarding people’s autonomy from the influence of widespread advertising procedures, especially in sensitive domains such as food or pharmaceutical advertising. Data Collection Information The data collection for the present experiments took place between March 2012 and February 2017 at the Universities of Heidelberg, Rotterdam, and Tübingen (study 1a: March 2012, Heidelberg; study 1b: September 2013, Heidelberg; study 2: October 2012, Heidelberg; study 3: January 2014, Tübingen; study 4: February 2017, Rotterdam; study 5: May/June 2014, Tübingen). Data collection was conducted by research assistants and bachelor students (study 5) under the supervision of the first author. Data were analyzed jointly by both authors. The MPT and power analyses were conducted by Mandy Hütter. APPENDIX A: DETAILS ON STUDY MATERIALS Instructions in Studies 1a, 1b, 2 Standard Condition In the upcoming impression formation phase, you are going to learn something about persons that are yet unknown to you. In the real world, one often gets correct information from third parties about other persons. Thus, the information that will be presented during the impression formation phase in terms of pleasant and unpleasant pictures will also be informative with regard to the person depicted in the respective photograph. Reversal Condition In the upcoming impression formation phase, you are going to learn something about persons that are yet unknown to you. In the real world, one often gets incorrect information from third parties about other persons. Thus, the information that will be presented during the impression formation phase in terms of pleasant and unpleasant pictures will be informative with regard to the person depicted in the respective photograph, but it will be downright opposite to what is actually correct. That means: pleasant pictures indicate that the respective person should be shown with unpleasant pictures, while unpleasant pictures indicate that the respective person should be shown with pleasant pictures. Instructions in Study 3 Research has shown that we quickly form impressions about unknown persons and that these evaluations are heavily influenced by the context in which we see a person. Specifically, research has shown that we come to like people who appear in positive contexts, and that we come to dislike people we see in negative contexts. Whereas the context may indeed be informative about the character of a person, there are many cases where directly applying the context evaluation to person impression formation would lead to an incorrect person evaluation. For example, an evil mafia boss can be encountered on a beautiful beach. Alternatively, very nice people might be encountered in criminal neighborhoods. In this research, we investigate whether people can apply contextual information when it is accurate and whether they can reverse the influence of contextual information when it is not. Therefore, in this study we will tell you when you can trust the contextual information you see people presented in, so that it can be directly applied to make appropriate person evaluations. We will also tell you when the contextual information is actually misleading, so that you should reverse the context evaluation to make the appropriate person evaluation. In the next phase, you will therefore be presented with many different persons appearing in positive or negative contexts. You will also be informed whether the context evaluation should be directly applied to the person evaluation, or whether it should be reversed when forming an impression of the person. After the impression formation phase, we want you to report for each person whether you like her or not. We are going to count how many of your evaluations were made properly—that is, applying the context evaluation directly when it is accurate, and reversing it when it is not. Participants in the financial incentives condition then received the following additional instructions: For each appropriate classification, you will receive €0.15 as a reward. This implies that if you get them all correct, you can gain up to €3.60 on top of your base compensation (€3.00) for participation in this study. Instructions in Study 4 Standard Condition New water brands are introduced to the market on a regular basis. Previous research on advertising shows that we quickly form impressions of unknown brands and that these impressions can be strongly influenced by the context in which we encounter a brand for the first time. Specifically, this means that we may learn to like brands that appear with positive images and learn to dislike brands that appear with negative images. In advertising, such images may very well provide information about a brand. For instance, marketers might invest more money in advertising for products that have proved successful in the past or are expected to sell well. Conversely, low-quality marketing might be reflective of a low-quality product. In this research, we investigate whether people are able to APPLY the positive or negative quality of the images to the brand. In the next phase, you will therefore be presented with different water brand logos appearing repeatedly with positive or negative images. Hence, you should start LIKING the brand when paired with a POSITIVE image. Conversely, you should start DISLIKING the brand when paired with a NEGATIVE image. Afterwards, we will ask you how you feel about the different water brands. SUMMARY: If you see a POSITIVE image, you should LIKE the brand. If you see a NEGATIVE image, you should DISLIKE the brand. Reversal Condition New water brands are introduced to the market on a regular basis. Previous research on advertising shows that we quickly form impressions of unknown brands and that these impressions can be strongly influenced by the context in which we encounter a brand for the first time. Specifically, this means that we may learn to like brands that appear with positive images and learn to dislike brands that appear with negative images. While such images may well provide information about a brand in advertising, many cases are conceivable in which the direct application of the positive or negative quality of an image would lead to erroneous impressions. For example, products may be of low quality despite positive advertising. Conversely, high-quality products can be put in a negative light by competitive brands. In this research, we investigate whether people are able to REVERSE the influence of the positive or negative quality of the images to the brand. In the next phase, you will therefore be presented with different water brand logos repeatedly appearing together with positive or negative images. Hence, you should start DISLIKING the brand when paired with a POSITIVE image. Conversely, you should start LIKING the brand when paired with a NEGATIVE image. Afterwards, we will ask you how you feel about the different water brands. SUMMARY: If you see a POSITIVE image, you should DISLIKE the brand. If you see a NEGATIVE image, you should LIKE the brand. Covert Condition New water brands are introduced to the market on a regular basis. This study investigates the spontaneous reactions of consumers to different water brand logos. Since you are unlikely to have seen these brands in stores before, we will first familiarize you with them. You will see a slideshow with images not only of these water brand logos but also of landscapes and people engaged in various activities. We hope this makes the slideshow more interesting to watch. Afterwards, we will ask you how you feel about the different water brands. Please pay careful attention to this slideshow. If you miss the presentations of some water brands, you will be less familiar with them than with others and this could affect the results. SUMMARY: Please pay careful attention to this slideshow. Instructions in Study 5 Previous research on advertising shows that we quickly form impressions of unknown brands and that these impressions can be strongly influenced by the context in which we encounter a brand for the first time. Specifically, this means that we may learn to like brands that appear in positive contexts and learn to dislike brands that we encounter in negative contexts. While the context may well provide information about a brand, many cases are conceivable in which the direct application of the context would lead to erroneous impressions. For example, products may be of low quality despite positive advertising. Conversely, high-quality products can be put in a negative light by competitive brands. In this research, we investigate whether people are able to apply contextual information when it is accurate and whether they can reverse the influence of contextual information when it is not. Therefore, in this study we will tell you when you can trust the contextual information you see brands presented in, so that it can be directly applied to make appropriate brand evaluations. We will also tell you when the contextual information is actually misleading, so that you should reverse the context evaluation to make the appropriate brand evaluation. In the next phase, you will therefore be presented with many different brand logos appearing in positive or negative contexts. You will also be informed whether the context evaluation should be directly applied to evaluate the brand, or whether it should be reversed when forming an impression of the brand. After the computer-based part, participants moved to an adjacent room in the lab where the brands presented in the previous part were lined up to be tasted. Participants received the following instructions: Now that you completed all tasks on the computer, we would like to give you an opportunity to refresh yourself. These cups are filled with water of the brands you were presented with in the computer experiment. It is up to you which water brands and how many of them you would like to taste. Please decide spontaneously. We thank Olivier Corneille, the editors, and three anonymous reviewers for providing helpful feedback on previous versions of this manuscript. This research was supported by an Emmy-Noether grant (HU 1978/4-1) and a Koselleck grant (FI 294/23-1) from the German Research Foundation. We also gratefully acknowledge financial support from the INSEAD R&D and the INSEAD Alumni Fund as well as data collection assistance offered by the Erasmus Behavioral Lab at Erasmus University. APPENDIX B: FREQUENCY DATA UNDERLYING MPT ANALYSIS Observed Frequencies in the Pleasant (+) and Unpleasant (–) Response Categories as a Function of US Valence and Instruction Condition and the Respective Manipulations Standard instructions Reversal instructions + – pc + – pc Σ observations Study 1a CS+ 143 73 .66 85 143 .37 CS– 61 155 .72 124 104 .46 888 Study 1b CS+ 86 34 .72 46 74 .38 CS– 27 93 .78 80 40 .67 480 Study 2 Low load CS+ 118 62 .66 70 98 .42 CS– 51 129 .72 101 67 .40 High load CS+ 106 74 .59 85 95 .47 CS– 81 99 .55 86 94 .52 1416 Study 3 Control CS+ 154 68 .69 81 141 .36 CS– 67 155 .70 139 83 .37 Incentives CS+ 162 30 .84 41 151 .21 CS– 34 158 .82 147 45 .23 1656 Study 4 CS+ 457 203 .69 351 357 .50 CS– 274 386 .58 327 381 .54 Covert instructions CS+ 404 280 .59 CS- 318 366 .54 4104 Study 5 CS+ 406 80 .84 163 323 .34 CS- 78 408 .84 297 189 .39 1944 Standard instructions Reversal instructions + – pc + – pc Σ observations Study 1a CS+ 143 73 .66 85 143 .37 CS– 61 155 .72 124 104 .46 888 Study 1b CS+ 86 34 .72 46 74 .38 CS– 27 93 .78 80 40 .67 480 Study 2 Low load CS+ 118 62 .66 70 98 .42 CS– 51 129 .72 101 67 .40 High load CS+ 106 74 .59 85 95 .47 CS– 81 99 .55 86 94 .52 1416 Study 3 Control CS+ 154 68 .69 81 141 .36 CS– 67 155 .70 139 83 .37 Incentives CS+ 162 30 .84 41 151 .21 CS– 34 158 .82 147 45 .23 1656 Study 4 CS+ 457 203 .69 351 357 .50 CS– 274 386 .58 327 381 .54 Covert instructions CS+ 404 280 .59 CS- 318 366 .54 4104 Study 5 CS+ 406 80 .84 163 323 .34 CS- 78 408 .84 297 189 .39 1944 NOTE.— CS+: CSs paired with positive USs; CS-: CSs paired with negative USs. The proportion of US valence-congruent responses (pc) varies with instruction condition. The difference in proportions between the standard and reversal conditions can be traced back to the controllable processes leading to a reversal of CS evaluations in the reversal condition. Standard instructions Reversal instructions + – pc + – pc Σ observations Study 1a CS+ 143 73 .66 85 143 .37 CS– 61 155 .72 124 104 .46 888 Study 1b CS+ 86 34 .72 46 74 .38 CS– 27 93 .78 80 40 .67 480 Study 2 Low load CS+ 118 62 .66 70 98 .42 CS– 51 129 .72 101 67 .40 High load CS+ 106 74 .59 85 95 .47 CS– 81 99 .55 86 94 .52 1416 Study 3 Control CS+ 154 68 .69 81 141 .36 CS– 67 155 .70 139 83 .37 Incentives CS+ 162 30 .84 41 151 .21 CS– 34 158 .82 147 45 .23 1656 Study 4 CS+ 457 203 .69 351 357 .50 CS– 274 386 .58 327 381 .54 Covert instructions CS+ 404 280 .59 CS- 318 366 .54 4104 Study 5 CS+ 406 80 .84 163 323 .34 CS- 78 408 .84 297 189 .39 1944 Standard instructions Reversal instructions + – pc + – pc Σ observations Study 1a CS+ 143 73 .66 85 143 .37 CS– 61 155 .72 124 104 .46 888 Study 1b CS+ 86 34 .72 46 74 .38 CS– 27 93 .78 80 40 .67 480 Study 2 Low load CS+ 118 62 .66 70 98 .42 CS– 51 129 .72 101 67 .40 High load CS+ 106 74 .59 85 95 .47 CS– 81 99 .55 86 94 .52 1416 Study 3 Control CS+ 154 68 .69 81 141 .36 CS– 67 155 .70 139 83 .37 Incentives CS+ 162 30 .84 41 151 .21 CS– 34 158 .82 147 45 .23 1656 Study 4 CS+ 457 203 .69 351 357 .50 CS– 274 386 .58 327 381 .54 Covert instructions CS+ 404 280 .59 CS- 318 366 .54 4104 Study 5 CS+ 406 80 .84 163 323 .34 CS- 78 408 .84 297 189 .39 1944 NOTE.— CS+: CSs paired with positive USs; CS-: CSs paired with negative USs. The proportion of US valence-congruent responses (pc) varies with instruction condition. The difference in proportions between the standard and reversal conditions can be traced back to the controllable processes leading to a reversal of CS evaluations in the reversal condition. REFERENCES Allen Chris T. , Madden Thomas J. ( 1985 ), “ A Closer Look at Classical Conditioning,” Journal of Consumer Research , 12 3 , 301 – 15 . Google Scholar CrossRef Search ADS Alter Adam L. , Oppenheimer Daniel M. , Epley Nicholas , Eyre Rebecca N. ( 2007 ), “Overcoming Intuition: Metacognitive Difficulty Activates Analytic Reasoning,” Journal of Experimental Psychology: General , 136 4 , 569 – 76 . Google Scholar CrossRef Search ADS PubMed Baeyens Frank , Eelen Paul , Crombez Geert , Vandenbergh Omer ( 1992 ), “ Human Evaluative Conditioning: Acquisition Trials, Presentation Schedule, Evaluative Style and Contingency Awareness,” Behaviour Research and Therapy , 30 2 , 133 – 42 . Google Scholar CrossRef Search ADS PubMed Bargh John A. ( 1994 ), “The Four Horsemen of Automaticity: Awareness, Intention, Efficiency, and Control in Social Cognition,” in Handbook of Social Cognition, Vol. 1: Basic Processes , 2nd ed. , ed. Wyer Robert S. Jr. , Srull Thomas K. , Hillsdale, NJ : Erlbaum , 1 – 40 . Batchelder William H. , Riefer David M. ( 1999 ), “ Theoretical and Empirical Review of Multinomial Process Tree Modeling,” Psychonomic Bulletin & Review , 6 1 , 57 – 86 . Google Scholar CrossRef Search ADS PubMed Biegler Paul , Vargas Patrick ( 2013 ), “ Ban the Sunset? Nonpropositional Content and Regulation of Pharmaceutical Advertising,” American Journal of Bioethics , 13 5 , 3 – 13 . Google Scholar CrossRef Search ADS PubMed Bierley Calvin , McSweeney Frances K. , Vannieuwkerk Renee ( 1985 ), “ Classical Conditioning of Preferences for Stimuli,” Journal of Consumer Research , 12 3 , 316 – 23 . Google Scholar CrossRef Search ADS Botvinick Matthew M. , Braver Todd S. , Barch Deanna M. , Carter Cameron S. , Cohen Jonathan D. ( 2001 ), “Conflict Monitoring and Cognitive Control,” Psychological Review , 108 3 , 624 – 52 . Google Scholar CrossRef Search ADS PubMed Cohen Jacob ( 1988 ), Statistical Power Analysis for the Behavioral Sciences , Hillsdale, NJ : Erlbaum . Corneille Olivier , Yzerbyt Vincent , Pleyers Gordy , Mussweiler Thomas ( 2009 ), “Beyond Awareness and Resources: Evaluative Conditioning May Be Sensitive to Processing Goals,” Journal of Experimental Social Psychology , 45 1 , 279 – 82 . Google Scholar CrossRef Search ADS Curran Tim , Hintzman Douglas L. ( 1995 ), “Violations of the Independence Assumption in Process Dissociation,” Journal of Experimental Psychology: Learning, Memory, and Cognition , 21 3 , 531 – 47 . Google Scholar CrossRef Search ADS PubMed Dalton Amy N. , Huang Li ( 2014 ), “ Motivated Forgetting in Response to Social Identity Threat,” Journal of Consumer Research , 40 6 , 1017 – 38 . Google Scholar CrossRef Search ADS De Houwer Jan ( 2006 ), “Using the Implicit Association Test Does Not Rule Out an Impact of Conscious Propositional Knowledge on Evaluative Conditioning,” Learning and Motivation , 37 2 , 176 – 87 . Google Scholar CrossRef Search ADS De Houwer Jan , Teige-Mocigemba Sarah , Spruyt Adriaan , Moors Agnes ( 2009 ), “ Implicit Measures: A Normative Analysis and Review,” Psychological Bulletin , 135 3 , 347 – 68 . Google Scholar CrossRef Search ADS PubMed De Houwer Jan , Thomas Sarah , Baeyens Frank ( 2001 ), “Associative Learning of Likes and Dislikes: A Review of 25 Years of Research on Human Evaluative Conditioning,” Psychological Bulletin , 127 6 , 853 – 69 . Google Scholar CrossRef Search ADS PubMed Dedonder Jonathan , Corneille Olivier , Yzerbyt Vincent , Kuppens Toon ( 2010 ), “Evaluative Conditioning of High-Novelty Stimuli Does Not Seem to Be Based on an Automatic Form of Associative Learning,” Journal of Experimental Social Psychology , 46 6 , 1118 – 21 . Google Scholar CrossRef Search ADS Deutsch Roland , Gawronski Bertram , Strack Fritz ( 2006 ), “ At the Boundaries of Automaticity: Negation as Reflective Operation,” Journal of Personality and Social Psychology , 91 3 , 385 – 405 . Google Scholar CrossRef Search ADS PubMed Deutsch Roland , Kordts-Freudinger Robert , Gawronski Bertram , Strack Fritz ( 2009 ), “Fast and Fragile: A New Look at the Automaticity of Negation Processing,” Experimental Psychology , 56 6 , 434 – 46 . Google Scholar CrossRef Search ADS PubMed Fazio Russell H. ( 1990 ), “Multiple Processes by Which Attitudes Guide Behavior: The MODE Model as an Integrative Framework,” in Experimental Social Psychology , Vol. 23 , ed. Zanna Mark P. , San Diego, CA : Academic Press , 75 – 109 . Fiedler Klaus , Hütter Mandy ( 2014 ), “The Limits of Automaticity,” in Dual-Process Theories of the Social Mind , ed. Sherman Jeffrey , Gawronski Bertram , Trope Yaacov , New York : Guilford , 497 – 513 . Fiedler Klaus , Unkelbach Christian ( 2011 ), “Evaluative Conditioning Depends on Higher Order Cognitive Processes,” Cognition & Emotion , 25 4 , 639 – 56 . Google Scholar CrossRef Search ADS PubMed Fitzsimons Gavan J. , Williams Patti ( 2000 ), “ Asking Questions Can Change Choice Behavior: Does It Do So Automatically or Effortfully?” Journal of Experimental Psychology: Applied , 6 3 , 195 – 206 . Google Scholar CrossRef Search ADS PubMed Gawronski Bertram , Balas Robert , Creighton Laura A. ( 2014 ), “ Can the Formation of Conditioned Attitudes Be Intentionally Controlled?” Personality and Social Psychology Bulletin , 40 4 , 419 – 32 . Google Scholar CrossRef Search ADS PubMed Gawronski Bertram , Bodenhausen Galen V. ( 2006 ), “ Associative and Propositional Processes in Evaluation: An Integrative Review of Implicit and Explicit Attitude Change,” Psychological Bulletin , 132 5 , 692 – 731 . Google Scholar CrossRef Search ADS PubMed Gawronski Bertram , Mitchell Derek G. , Balas Robert ( 2015 ), “ Is Evaluative Conditioning Really Uncontrollable? A Comparative Test of Three Emotion-Focused Strategies to Prevent the Acquisition of Conditioned Preferences,” Emotion , 15 5 , 556 – 68 . Google Scholar CrossRef Search ADS PubMed Gibson Bryan ( 2008 ), “ Can Evaluative Conditioning Change Attitudes Toward Mature Brands? New Evidence from the Implicit Association Test,” Journal of Consumer Research , 35 1 , 178 – 88 . Google Scholar CrossRef Search ADS Gorn Gerald J. ( 1982 ), “ The Effects of Music in Advertising on Choice Behavior: A Classical Conditioning Approach,” Journal of Marketing , 46 1 , 94 – 101 . Google Scholar CrossRef Search ADS Hofmann Wilhelm , De Houwer Jan , Perugini Marco , Baeyens Frank , Crombez Geert ( 2010 ), “ Evaluative Conditioning in Humans: A Meta-Analysis,” Psychological Bulletin , 136 3 , 390 – 421 . Google Scholar CrossRef Search ADS PubMed Hu Xiangen , Batchelder William H. ( 1994 ), “The Statistical Analysis of Engineering Processing Tree Models with the EM Algorithm,” Psychometrika , 59 1 , 21 – 47 . Google Scholar CrossRef Search ADS Hütter Mandy , Klauer Karl C. ( 2016 ), “Applying Processing Trees in Social Psychology,” European Review of Social Psychology , 27 1 , 116 – 59 . Google Scholar CrossRef Search ADS Hütter Mandy , Sweldens Steven ( 2013 ), “Implicit Misattribution of Evaluative Responses: Contingency-Unaware Evaluative Conditioning Requires Simultaneous Stimulus Presentations,” Journal of Experimental Psychology: General , 142 3 , 638 – 43 . Google Scholar CrossRef Search ADS PubMed Hütter Mandy , Sweldens Steven , Stahl Christoph , Unkelbach Christian , Klauer Karl C. ( 2012 ), “Dissociating Contingency Awareness and Conditioned Attitudes: Evidence of Contingency-Unaware Evaluative Conditioning,” Journal of Experimental Psychology: General , 141 3 , 539 – 57 . Google Scholar CrossRef Search ADS PubMed Ito Tiffany A. , Larsen Jeff T. , Smith N. Kyle , Cacioppo John T. ( 1998 ), “ Negative Information Weighs More Heavily on the Brain: The Negativity Bias in Evaluative Categorizations,” Journal of Personality and Social Psychology , 75 4 , 887 – 900 . Google Scholar CrossRef Search ADS PubMed Jacoby Larry L. ( 1991 ), “A Process Dissociation Framework—Separating Automatic from Intentional Uses of Memory,” Journal of Memory and Language , 30 5 , 513 – 41 . Google Scholar CrossRef Search ADS Jacoby Larry L. , Shrout Patrick E. ( 1997 ), “Toward a Psychometric Analysis of Violations of the Independence Assumption in Process Dissociation,” Journal of Experimental Psychology: Learning, Memory, and Cognition , 23 2 , 505 – 10 . Google Scholar CrossRef Search ADS Janiszewski Chris , Warlop Luk ( 1993 ), “ The Influence of Classical Conditioning Procedures on Subsequent Attention to the Conditioned Brand,” Journal of Consumer Research , 20 2 , 171 – 89 . Google Scholar CrossRef Search ADS Johar Gita V. , Maheswaran Durairaj , Peracchio Laura A. ( 2006 ), “Mapping the Frontiers: Theoretical Advances in Consumer Research on Memory, Affect, and Persuasion,” Journal of Consumer Research , 33 1 , 139 – 49 . Google Scholar CrossRef Search ADS Jones Christopher R. , Fazio Russell H. , Olson Michael A. ( 2009 ), “Implicit Misattribution as a Mechanism Underlying Evaluative Conditioning,” Journal of Personality and Social Psychology , 96 5 , 933 – 48 . Google Scholar CrossRef Search ADS PubMed Klauer Karl Christoph , Dittrich Kerstin , Scholtes Christine , Voss Andreas ( 2015 ), “The Invariance Assumption in Process-Dissociation Models: An Evaluation Across Three Domains,” Journal of Experimental Psychology: General , 144 1 , 198 – 221 . Google Scholar CrossRef Search ADS PubMed Lang Peter J. , Bradley Margaret M. , Cuthbert Bruce N. ( 2008 ), International Affective Picture System (IAPS): Affective Ratings of Pictures and Instruction Manual . Technical Report A-8, Gainesville : University of Florida . Mierop Adrien , Hütter Mandy , Corneille Olivier ( 2017 ), “Resource Availability and Explicit Memory Largely Determine Evaluative Conditioning Effects in a Paradigm Claimed to Be Conducive to Implicit Attitudes Acquisition,” Social Psychological and Personality Science , 8 7 , 758 – 67 . Google Scholar CrossRef Search ADS Mitchell Chris J. , De Houwer Jan , Lovibond Peter F. ( 2009 ), “ The Propositional Nature of Human Associative Learning,” Behavioral and Brain Sciences , 32 2 , 183 – 98 . Google Scholar CrossRef Search ADS PubMed Moors Agnes , De Houwer Jan ( 2006 ), “Automaticity: A Conceptual and Theoretical Analysis,” Psychological Bulletin , 132 2 , 297 – 326 . Google Scholar CrossRef Search ADS PubMed Moshagen Morten ( 2010 ), “MultiTree: A Computer Program for the Analysis of Multinomial Processing Tree Models,” Behavior Research Methods , 42 1 , 42 – 54 . Google Scholar CrossRef Search ADS PubMed Olson Michael A. , Fazio Russell H. ( 2001 ), “ Implicit Attitude Formation through Classical Conditioning,” Psychological Science , 12 5 , 413 – 7 . Google Scholar CrossRef Search ADS PubMed Page Monte M. ( 1969 ), “Social Psychology of a Classical Conditioning of Attitudes Experiment,” Journal of Personality and Social Psychology , 11 2 , 177 – 86 . Google Scholar CrossRef Search ADS Payne B. Keith , Bishara Anthony J. ( 2009 ), “An Integrative Review of Process Dissociation and Related Models in Social Cognition,” European Review of Social Psychology , 20 1 , 272 – 314 . Google Scholar CrossRef Search ADS Pham Michel T. , Johar Gita V. ( 1997 ), “ Contingent Processes of Source Identification,” Journal of Consumer Research , 24 3 , 249 – 65 . Google Scholar CrossRef Search ADS Pleyers Gordy , Corneille Olivier , Yzerbyt Vincent , Luminet Olivier ( 2009 ), “Evaluative Conditioning May Incur Attentional Costs,” Journal of Experimental Psychology: Animal Behavior Processes , 35 2 , 279 – 85 . Google Scholar CrossRef Search ADS PubMed Rogers Robert D. , Monsell Stephen ( 1995 ), “Costs of a Predictable Switch between Simple Cognitive Tasks,” Journal of Experimental Psychology: General , 124 , 207 – 31 . Google Scholar CrossRef Search ADS Shanks David R. ( 2010 ), “ Learning: From Association to Action,” Annual Review of Psychology , 61 , 273 – 301 . Google Scholar CrossRef Search ADS PubMed Shapiro Stewart ( 1999 ), “When an Ad’s Influence Is Beyond Our Conscious Control: Perceptual and Conceptual Fluency Effects Caused by Incidental Ad Exposure,” Journal of Consumer Research , 26 1 , 16 – 36 . Google Scholar CrossRef Search ADS Shimp Terence A. , Stuart Elnora W. , Engle Randall W. ( 1991 ), “ A Program of Classical-Conditioning Experiments Testing Variations in the Conditioned-Stimulus and Context,” Journal of Consumer Research , 18 1 , 1 – 12 . Google Scholar CrossRef Search ADS Stahl Christoph , Klauer Karl C. ( 2007 ), “HMMTree: A Computer Program for Latent-Class Hierarchical Multinomial Processing Tree Models,” Behavior Research Methods , 39 2 , 267 – 73 . Google Scholar CrossRef Search ADS PubMed Stuart Elnora W. , Shimp Terence A. , Engle Randall W. ( 1987 ), “ Classical Conditioning of Consumer Attitudes: Four Experiments in an Advertising Context,” Journal of Consumer Research , 14 3 , 334 – 49 . Google Scholar CrossRef Search ADS Sweldens Steven , Corneille Olivier , Yzerbyt Vincent ( 2014 ), “ The Role of Awareness in Attitude Formation through Evaluative Conditioning,” Personality and Social Psychology Review , 18 2 , 187 – 209 . Google Scholar CrossRef Search ADS PubMed Sweldens Steven , Tuk Mirjam , Hütter Mandy ( 2017 ), “How to Study Consciousness in Consumer Research: A Commentary on Williams and Poehlman,” Journal of Consumer Research , 44 2 , 266 – 75 . Sweldens Steven , van Osselaer Stijn M. J. , Janiszewski Chris ( 2010 ), “ Evaluative Conditioning Procedures and the Resilience of Conditioned Brand Attitudes,” Journal of Consumer Research , 37 3 , 473 – 89 . Google Scholar CrossRef Search ADS Tukey John W. ( 1977 ), Exploratory Data Analysis, Reading, PA : Addison-Wesley . Wegner Daniel M. ( 1994 ), “Ironic Processes of Mental Control,” Psychological Review , 101 1 , 34 – 52 . Google Scholar CrossRef Search ADS PubMed Whitmer Anson J. , Banich Marie T. ( 2007 ), “Inhibition versus Switching Deficits in Different Forms of Rumination,” Psychological Science , 18 6 , 546 – 53 . Google Scholar CrossRef Search ADS PubMed Williams Lawrence E. , Poehlman Andrew ( 2017 ), “Conceptualizing Consciousness in Consumer Research,” Journal of Consumer Research , 44 2 , 231 – 51 . © The Author(s) 2018. Published by Oxford University Press on behalf of Journal of Consumer Research, Inc. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Consumer Research Oxford University Press

Dissociating Controllable and Uncontrollable Effects of Affective Stimuli on Attitudes and Consumption

Loading next page...
 
/lp/ou_press/dissociating-controllable-and-uncontrollable-effects-of-affective-Uc0Fcmbs7R
Publisher
University of Chicago Press
Copyright
© The Author(s) 2018. Published by Oxford University Press on behalf of Journal of Consumer Research, Inc. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com
ISSN
0093-5301
eISSN
1537-5277
D.O.I.
10.1093/jcr/ucx124
Publisher site
See Article on Publisher Site

Abstract

Abstract This research studies a fundamental and seemingly straightforward question: Can basic advertising elements, such as the presence of attractive imagery, have uncontrollable effects on consumers’ attitudes and consumption decisions? Answering this question is methodologically challenging, because the presence of an uncontrollable process can be masked by a simultaneously operating controllable process. We argue first that existing methods conflate the contribution of both processes and are therefore unable to measure the presence of an uncontrollable process reliably. To solve the conundrum, we present a novel application of processing tree modeling. Evaluative conditioning is employed as a paradigm to study the influence of affective visual stimuli on attitudes and behavior. Across six experiments, we demonstrate the validity of the model parameters estimating controllable and uncontrollable processes. As predicted, the parameter estimate of the controllable process is susceptible to cognitive resources and levels of motivation to exert control. The parameter estimate of the uncontrollable process appears unaffected by these factors. We also demonstrate the external validity of our findings and their relevance to stimuli and instructions typical for consumer research. Finally, we find that controllable and uncontrollable processes are both predictive of product choice and consumption. We discuss implications for consumer protection. attitudes, evaluative conditioning, evaluative learning, controllability, automaticity, multinomial processing tree models The evaluative conditioning (EC) procedure is an experimental analog to the acquisition of consumer attitudes in everyday life. In a typical EC procedure, a conditioned stimulus (CS; e.g., a brand name) repeatedly co-occurs in close spatiotemporal proximity with a valenced unconditioned stimulus (US; e.g., a beautiful image or a celebrity endorser). As a result, evaluations of the CS typically change in the direction of the valence of the US—a phenomenon known as the EC effect. This basic procedure is considered a key representation of evaluative learning in general (for reviews, see De Houwer, Thomas, and Baeyens 2001; Hofmann et al. 2010). EC also serves as the prototypical paradigm to study advertising effects on brand attitudes (Allen and Madden 1985; Bierley, McSweeney, and Vannieuwkerk 1985; Gibson 2008; Gorn 1982; Janiszewski and Warlop 1993; Shimp, Stuart, and Engle 1991; Stuart, Shimp, and Engle 1987; Sweldens, Van Osselaer, and Janiszewski 2010). In the present research, we propose and validate a novel methodological approach to assess whether and to what degree EC procedures can have uncontrollable effects on attitudes and consumption. As we will outline, we study controllability first, because it can play a crucial role in a theoretical debate that has been going on for decades: Are the processes underlying EC effects characterized by features of automaticity? The second reason is substantive in nature—such uncontrollable effects would have important implications for consumer welfare and protection. INTRODUCTION In the introduction to this article, we first highlight the theoretical relevance of studying controllability. Next, we illuminate the methodological approach we propose, introducing a multinomial processing tree model to separate controllable from uncontrollable processes in EC. We then propose and present a series of studies to test and validate this model. We return to the substantive implications for consumer protection in the General Discussion. Automaticity, Controllability, and EC The possible contribution of automatically operating learning processes has a long and contentious history in EC. Consistent with Williams and Poehlman’s (2017) recent treatise on consciousness, much of this debate has focused on whether EC effects can be established without awareness of the contingencies between the stimuli (Sweldens, Corneille, and Yzerbyt 2014). However, the concept of automaticity is defined by several features that do not necessarily co-occur (Bargh 1994; Williams and Poehlman 2017). Instead, cognitive processes can be characterized by automaticity in varying configurations. Therefore, it is unwarranted to infer the presence or absence of other automaticity features from evidence regarding awareness (Fiedler and Hütter 2014; Moors and De Houwer 2006; Sweldens, Tuk, and Hütter 2017). Notably, research investigating automaticity features other than awareness often indicates that EC would not be automatic in these respects. For example, three recent articles investigating EC’s efficiency found that EC effects do not occur when participants are under cognitive load (Dedonder et al. 2010; Mierop, Hütter, and Corneille 2017; Pleyers et al. 2009). Similarly, research indicates that EC effects are influenced by intentionality or participants’ processing goals during the EC procedure (Corneille et al. 2009; Fiedler and Unkelbach 2011). Studying the controllability of the learning process in EC has the potential to advance the debate on this front significantly. Current State of the Art: Explicit Evaluations Can, but Implicit Evaluations Cannot, Be Controlled We know of only two previous studies investigating the controllability of EC effects. Gawronski, Balas, and Creighton (2014) instructed participants to prevent or promote the effect of the US pairings on their CS evaluations in the conditioning procedure. The authors observed a dissociation between an indirect measure of evaluation (an evaluative priming task) and a direct self-report measure. While the indirect measure showed standard, unmodulated EC effects in all conditions, the direct measure appeared subject to expressive control. The same pattern of results was obtained in subsequent work comparing different strategies of exerting control. Irrespective of the strategy participants employed (e.g., suppressing emotional reactions to the US or reappraising the valence of the US), it appeared that indirect measures of evaluation were not subject to control, but deliberate responses on direct measures appeared controllable (Gawronski, Mitchell, and Balas 2015). In summary, this research suggests that, while the associative structure might change (as revealed in the indirect measures of evaluation), consumers retain the ability to control their explicit evaluations. Why We Investigate the Uncontrollability of Direct Evaluations There are two important reasons why the current research focuses on direct rather than indirect measures: a methodological and a substantive one. First, whereas indirect measures of evaluation are often good reflections of associative structure, they may not give a complete picture of what people have learned. This issue is especially acute in contexts where negations come into play. For instance, Deutsch, Gawronski, and Strack (2006) presented their participants with affirmed or negated primes in an evaluative priming task. In indirect measures of evaluation, attitudes toward primes are inferred from participants’ response times in classifying subsequent target stimuli. The presentation of the prime “pleasure” facilitated responses for positive targets and inhibited responses for negative targets. However, Deutsch and colleagues (2006) found the same pattern of response facilitation and inhibition when the prime “no pleasure” was presented. Hence, a negated representation still improved reaction times for targets of the opposite valence. These results illustrate why it is problematic to rely on indirect measures to study the controllability of EC. Consider the case of a controllability experiment in which participants aim to reverse or negate the effects of US valence on CS evaluations. If EC effects would be completely controllable, participants might encode CSs that were shown with positive USs as associated with “not positive” instead. However, when this CS is later encountered as a prime in an evaluative priming task, its “not positive” association can still facilitate response times to positive targets and thereby indicate a regular EC effect, even though full control was exerted over learning and participants actually hold an association of opposite valence. Hence, reaction-time-based evaluative measures contain the risk of considerably overestimating the contribution of uncontrollable learning, which should be avoided in a research context where the very existence of a phenomenon (in this case, uncontrollability) is to be demonstrated. Second, from a substantive viewpoint it is similarly important to study direct evaluations. As indicated above, according to the current state of the literature, EC can have uncontrollable effects on indirect measures of evaluation, implying that conditioning procedures can change concepts’ associative structure uncontrollably (Gawronski et al. 2014; Gawronski et al. 2015). At the same time, the current evidence indicates that control can still be exerted at the expression stage, when consumers validate the output of the associative system (i.e., in their direct evaluations). Substantively, one could argue that it does not matter at which stage control is being exerted. As long as consumers are successful at negating the effects of EC (e.g., by invalidating the output of the associative system), their autonomy has not been compromised as they retain control over their deliberate behavior. If, however, EC would have uncontrollable effects on direct evaluations, this would be more problematic from a consumer protection viewpoint. After all, this would indicate that consumers’ control strategies failed at both the encoding and the validation stages, opening the door for uncontrollable effects on deliberate behavior as well. Why We Need Better Measurement Methods The main methodological challenge in studying the uncontrollability of EC stems from the difficulty of distinguishing the contribution of multiple processes in any measurement outcome. The presence of uncontrollable effects needs to be detected in addition to the controllable effects of EC. This problem originates in the fact that no measure of evaluation is process-pure: both direct and indirect measures of evaluation can be reflective of both explicitly and implicitly acquired evaluations (De Houwer 2006; De Houwer et al. 2009). To illustrate this problem in the current context, consider the following, hypothetical example of a researcher who aims to study the controllability of EC. The researcher designs an EC experiment featuring a conditioning procedure in which several CSs are paired with positive USs and other CSs are paired with negative USs. To study the controllability of the process, the researcher implements three experimental conditions. In the control condition, participants receive no special instructions as they go through the conditioning procedure and subsequently provide their evaluative ratings of the CS on a measure of evaluation (whether the measure is direct or indirect does not matter for this example). In the prevent condition, participants are asked to prevent any influence the conditioning procedure might have on their evaluative ratings of the CSs. In the reverse condition, participants are asked to reverse any influence the conditioning procedure might have—that is, to evaluate CSs paired with positive images more negatively than those paired with negative images. Consider the hypothetical pattern of results displayed in figure 1. FIGURE 1 View largeDownload slide HYPOTHETICAL RESULTS OF AN EXPERIMENT ON CONTROLLABILITY IN EC NOTE.— CSs– = Negatively paired conditioned stimuli, CSs+ = positively paired conditioned stimuli. FIGURE 1 View largeDownload slide HYPOTHETICAL RESULTS OF AN EXPERIMENT ON CONTROLLABILITY IN EC NOTE.— CSs– = Negatively paired conditioned stimuli, CSs+ = positively paired conditioned stimuli. At first glance, the researcher could be tempted to conclude that the EC effect is completely under participants’ control. After all, consistent with the hypothesis that the effect is controllable, the researcher observes a standard EC effect in the control condition, no EC effect in the prevent condition, and a reversed EC effect in the reverse condition. However, a more careful look at the results reveals the limitations of the method. First, the null result in the prevent condition is, like most null results, difficult to interpret. Even if EC had an uncontrollable effect on attitudes, participants in this condition could still easily comply with the task instructions to prevent the influence of the conditioning procedure on their evaluative ratings by employing several response strategies (e.g., by clicking the midpoint of the scale in the case of direct measures, or by responding randomly in the case of either direct or indirect measures of evaluation). Second, the reversed EC effect in the reverse condition is also ambiguous. In this condition, the controllable process may stand in opposition to an uncontrollable one. In such cases, one should account for the possibility that the uncontrollable process might generate a smaller (but non-negligible) effect than the controllable process. In an experimental setting with many CSs, control might, for example, fail for a (statistically significant) minority of CSs, whereas for the majority of CSs control succeeds and a reversed EC effect is observed. This would be consistent with work studying other features of automaticity in EC. For example, research investigating EC’s dependence on awareness of the CS-US contingency found a much larger proportion of CSs showing a conditioning effect characterized by contingency memory than CSs showing a conditioning effect without contingency memory (Hütter et al. 2012; Hütter and Sweldens 2013). The researcher’s problem is that an aggregated measure of evaluation does not distinguish between a situation in which participants can reverse evaluations for some CSs but not for others, and a competing possibility in which participants are in fact able to reverse the EC effect for all CSs but form less extreme evaluations. Whereas the former possibility would be consistent with the existence of an uncontrollable process in EC, the latter possibility would imply that EC is fully controllable. Overcoming the problems caused by aggregation requires the use of the processing tree framework to distinguish the contributions of different processes on the level of trials in a single measure. THE PROCESSING TREE FRAMEWORK Processing tree models are a class of statistical models that allow for separating and quantifying the effects of processes underlying performance in a single task (Batchelder and Riefer 1999). In the present research, we develop a multinomial processing tree (MPT) model to measure the contribution of controllable and uncontrollable processes to the formation of evaluations in a direct measure. Processing tree models have in common that latent cognitive processes are related to observable responses. The contribution of each process is estimated in terms of a parameter that expresses the probability of that process operating. Processing tree models can be applied to categorical data (e.g., “pleasant” and “unpleasant” responses) with the restriction that the database needs to be rich enough to allow for the separation of multiple processes. Therefore, in many cases processing trees require the use of different experimental conditions that lead to distinct relations of the latent cognitive processes to the observable responses. This framework has been applied in a variety of contexts (Hütter and Klauer 2016; Payne and Bishara 2009). The present work examines whether this framework is suitable for investigating the controllability of EC effects. That is, we develop and validate an MPT model to separate controllable and uncontrollable attitude acquisition processes in an EC paradigm. To dissociate the processes contributing (i.e., the latent constructs) to EC effects (i.e., the observable responses), we create a standard and a reversal version of the conditioning phase following the process dissociation logic introduced by Jacoby (1991). In the standard condition, participants are told that they can apply the valence of the USs to form an evaluation of the CSs. In this condition, both controllable and uncontrollable processes lead to regular EC effects. In the reversal condition, the instructions state that the US valence should be opposite to the one presented for participants to form accurate evaluations of the CS. To the extent that US valence has uncontrollable effects on CS attitudes, the CS will still acquire the unmodified US valence. Hence, in the reversal condition, controllable and uncontrollable processes have opposite effects on CS evaluations. These predictions can be visualized in terms of the processing tree model depicted in figure 2. The model relates contributions of controllable valence transfer (estimated by the c-parameter) and uncontrollable valence transfer (the u-parameter) to dichotomous evaluative responses while controlling for unsystematic processes (i.e., that do not vary systematically with US valence and instructions; the r-parameter). FIGURE 2 View largeDownload slide PROCESSING TREE MODEL OF POSITIVE (+) AND NEGATIVE (–) EVALUATIONS FORMED IN THE CONDITIONING PHASE IN THE STANDARD AND REVERSAL CONDITIONS FOR A POSITIVELY (CS+) AND NEGATIVELY (CS–) PAIRED CS NOTE.—The rectangles on the left denote the stimuli, while the rectangles on the right denote the responses. The branches of the processing tree represent the combination of cognitive processes postulated by the model. c = probability of acquiring an evaluation via controllable processes; u = probability of acquiring an evaluation via uncontrollable processes given the absence of controllable processes; r = positive response tendency. FIGURE 2 View largeDownload slide PROCESSING TREE MODEL OF POSITIVE (+) AND NEGATIVE (–) EVALUATIONS FORMED IN THE CONDITIONING PHASE IN THE STANDARD AND REVERSAL CONDITIONS FOR A POSITIVELY (CS+) AND NEGATIVELY (CS–) PAIRED CS NOTE.—The rectangles on the left denote the stimuli, while the rectangles on the right denote the responses. The branches of the processing tree represent the combination of cognitive processes postulated by the model. c = probability of acquiring an evaluation via controllable processes; u = probability of acquiring an evaluation via uncontrollable processes given the absence of controllable processes; r = positive response tendency. Several explanatory notes are in order. First, the main theoretical interest of our research lies not in demonstrating instances of failure of control per se (1 – c), but in whether there is an uncontrollable process (u) contributing to evaluations if control fails (this does not necessarily follow from the former). To elaborate, the model perfectly allows for trials or instances where people do not exert control or follow instructions—that is, where they do not apply US valence to the CS evaluation as instructed (1 – c). There are plenty of reasons why people might not exert control that are not due to an uncontrollable effect of the affective stimuli. For instance, they might not be motivated to follow instructions or they might not be paying attention to the stimuli, which would all lower the c-parameter (and hence increase 1 – c). It is important to realize, however, that none of those circumstances on their own would necessarily increase the u-parameter, because in all those circumstances it remains possible that no valence is transferred from the US to the CS. The u-parameter is derived only from those cases (across conditions) where US valence is applied irrespective of the instructions given to the participant. Therefore, the u-parameter strictly measures the uncontrolled application of US valence. Consequently, in a second step it still needs to be demonstrated that such instances of uncontrolled US valence transfer are characterized by features of automaticity, before one regards the u-parameter as a reflection of uncontrollable (i.e., automatic) valence transfer. Second, it is important to note that because the MPT model is applied to dichotomous response options, it is not necessary to imagine a stimulus or generate an affective reaction that is equally strong as the actual US presented for control to be registered by the model. The quality (i.e., positive or negative valence), but not the quantity (i.e., intensity), is critical in producing the c-parameter. Third, the MPT model is specifically designed to test for the presence of uncontrollable processes over and above controllable ones, without making any assumptions about the sequence of processes operating during encoding. That is, the model should neither imply that uncontrollable processes would operate only when control fails nor deny the possibility that uncontrollable processes take precedence over controllable ones during encoding. The model constitutes a mere measurement model that assesses whether there is a contribution of uncontrollable processes (u-parameter) to attitude acquisition after controllable processes (c-parameter) and unsystematic processes (r-parameter) have been partialed out. Finally, the explicit nature of the instructions has several consequences that are important for theorizing on automatic attitude acquisition. Specifically, the instructions used in this controllability setting are quite explicit about the CS-US pairings and their expected consequences for CS attitudes. Hence, awareness levels can be expected to be high in the present research. High levels of awareness, however, do not guarantee high levels of controllability. To the contrary, participants may be aware of the stimuli, their pairing, and their evaluative reaction to the US, but not be able to reverse the effect of US valence on the evaluation of the CS. Experiments in the domain of thought control illustrate this fact. Even though individuals are fully aware of unwanted thoughts entering consciousness, these thoughts are difficult to control. In fact, attempting to exert control increases the probability of unwanted thoughts occurring in consciousness (Wegner 1994; Whitmer and Banich 2007). Validating the Model and Its Parameters in a Series of Studies Developing and applying an MPT model in a new research context does not by itself guarantee the reliability and validity of the parameters obtained (Hütter and Klauer 2016). We follow a multistudy approach to demonstrate the method’s ability to distinguish controllable from uncontrollable processes in EC, reflected by its c- and u-parameters. This requires demonstrating several desired characteristics of the model and parameter estimates. The first indicator is model fit; that is, in all studies we assess whether the model we developed constitutes a suitable description of the data obtained. Fit statistics assess the degree to which the predicted response frequencies deviate from the observed response frequencies on the categorical evaluative measure across experimental conditions. The asymptotically χ2-distributed log-likelihood statistic G² is used to assess the goodness of fit of MPT models (Hu and Batchelder 1994). A good fit (i.e., low deviation) would be expected when the model can accommodate the observed category frequencies. The goodness-of-fit statistic also aids hypothesis testing by comparing different specifications of the MPT model. Specifically, the statistic ΔG² assesses the change in fit when parameters are restricted to a specific value (i.e., zero) or equal to another. Setting parameters to zero and observing the effect on model fit is a way to test whether a parameter is significantly greater than zero (i.e., it explains a substantial proportion of the data) and must be retained in the model. Conducting this test on the uncontrollability (u) parameter is of primary theoretical interest in the present line of research. Model fit alone, however, does not guarantee that the model parameters measure the constructs they are intended to measure (Hütter and Klauer 2016). Hence, the second major challenge is to demonstrate the construct validity of the model parameters, relating to the crucial question of whether the c- and u-parameters indeed capture processes that differ in their degree of automaticity (Bargh 1994). A particularly difficult question is whether the process captured by the u-parameter does indeed reflect uncontrollable effects of US valence on CS attitudes (i.e., it demonstrates automaticity). This is not guaranteed by the MPT parameters per se: a significant u-parameter merely indicates the presence of a statistically significant number of CSs where control was not applied, yet a (non-reversed) EC effect was acquired. This does not necessarily imply automaticity; perhaps control could have been exerted if participants had been more motivated or had more opportunity for processing. To demonstrate the automaticity of a process, one needs to demonstrate that it bypasses manipulations of processing capacity or motivation to control (Fazio 1990). In contrast, the controllability of a process as represented by the c-parameter should depend on whether participants are motivated to exert control and whether they have the cognitive resources to do so. Studies 2 and 3 therefore test the differential susceptibility of the c- and u-parameters to manipulations of cognitive capacity and motivation, establishing their construct validity. We will also investigate the external validity of the method—that is, whether the method and findings are relevant to more than one specific paradigm. We thus investigate the current questions in paradigms that employ faces (studies 1a, 1b, 2, 3) and consumer brands (studies 4, 5) as CSs. Moreover, we also investigate the relevance of our approach for the broader EC literature. Because the main theoretical interest applies to the question of whether visual affective stimuli can have uncontrollable effects, we normally implore participants to exert control to the best of their ability. This is, however, atypical for EC research, where the purpose of the stimulus pairings is generally concealed. Therefore, in study 4 we will create a context that allows us to assess the relevance of our conclusions for more typical EC paradigms. Equally important is to demonstrate the predictive validity of the model parameter estimates—that is, to demonstrate that the model’s parameter estimates can predict scores on other criterion variables. A first important question is whether the MPT parameter estimates correlate as predicted with independently assessed attitude measures. Specifically, the model’s c-parameter, which measures the controllable influence of US valence, should correlate positively with the EC effect in the standard condition, but correlate inversely in a condition in which participants aim to reverse the influence of US valence. Conversely, the model’s u-parameter, which measures the uncontrollable influence of US valence, should correlate positively with the EC effect in both the standard and reversal conditions, as this process should not be affected by instructions or participants’ efforts to exert control. One key advantage of within-participant designs is that parameter estimates are obtained at the individual level (between-participants designs provide only aggregated parameter estimates per condition). Therefore, predictive validity of this kind can be assessed in studies featuring within-participant designs (studies 1b, 3, and 5). The second aspect of predictive validity, and from a substantive point of view arguably the most important one, is whether the model parameter estimates also predict actual product choice and consumption. Investigating the predictive validity of the parameters for consumers’ brand choice and consumption is the main goal of study 5. STUDY 1A: A BETWEEN-PARTICIPANTS TEST OF THE MPT MODEL By comparing responses in the standard and reversal conditions, study 1a investigates whether there is a significant number of CSs for which participants are unable to reverse the US valence when required. The standard and reversal instruction conditions are implemented between-participants, implying that participants need to reverse either no or all US valence information. Method Participants In this first study, 37 students (25 female) of different majors at the University of Heidelberg took part (Mage = 22.35, SDage = 4.81). Participants chose between receiving course credit or €3.00 (approximately US$4.00) as a reward for this study, which took about 25 minutes. Design The experiment employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: preratings vs. postratings) × 2 (instruction condition: standard vs. reversal) mixed design with repeated measures on the first two factors. Materials A set of 102 black-and-white pictures of human faces was used as the CS repertory (Hütter et al. 2012). For each participant, CSs were selected from that pool based on an initial evaluative rating. Fifty unambiguously pleasant and 50 unambiguously unpleasant pictures from the International Affective Picture System (IAPS; Lang, Bradley, and Cuthbert 2008) served as USs. Pleasant and unpleasant pictures differed in valence, t(98) = 60.77, p < .001. They also differed in dominance, t(98) = 19.66, p < .001 (unpleasant pictures were relatively more dominant), but not in arousal, t(98) = –.02, p = .98. Procedure Participants first rated the valence of the 102 portraits on a continuous scale with the endpoints “very unpleasant” (–100) and “very pleasant” (100). For each participant, the 24 portraits with the most neutral rating were selected as CSs. In the subsequent conditioning phase, the CSs were paired with the USs. Each CS was randomly assigned six different USs. Each CS was shown once with each assigned US in a simultaneous presentation. Pairings were presented for 3,000 milliseconds with an intertrial interval of 200 milliseconds. In the standard condition, participants were told that the US valence is informative regarding the person (CS). In the reversal condition, instructions also stated that US valence should be opposite from what is presented. Hence, when an unpleasant picture is presented it should actually be pleasant, and vice versa (precise instructions are in appendix A). After the conditioning phase, participants indicated for every CS whether their evaluation of it was “pleasant” or “unpleasant.” This is the key dichotomous measure to which the MPT model can be applied. Additionally, we obtained continuous evaluative ratings of the 24 CSs a second time (on the same scale as the preratings). After the completion of the experiment, participants were thanked and dismissed. Results Evaluative Ratings The unstandardized continuous evaluative ratings are presented for the two instruction conditions in figure 3. They were analyzed with a 2 (US valence) × 2 (time of evaluative rating) × 2 (instruction condition) mixed ANOVA. EC effects in this design are revealed by the interaction between the US valence and time of evaluative ratings factors, as the EC procedure should affect CS evaluations dependent on US valence. There was no overall EC effect, as the two-way interaction between US valence and time of measurement was not significant (F(1, 35) = 2.35, p = .13, ηp2 = .06). Instead, the EC effect was moderated by instruction condition (F(1, 35) = 21.86, p < .001, ηp2 = .38). Additionally, the overall non-significant main effect of US valence (F(1, 35) = 1.68, p = .20, ηp2 = .05) was dependent on instruction condition (F(1, 35) = 17.84, p < .001, ηp2 = .34). The remaining effects were non-significant (smallest p = .47). FIGURE 3 View largeDownload slide EVALUATIVE RATINGS IN STUDY 1A BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. FIGURE 3 View largeDownload slide EVALUATIVE RATINGS IN STUDY 1A BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. The interaction of US valence and time of measurement was assessed separately for the instruction conditions, revealing a significant EC effect in the standard condition (F(1, 35) = 18.76, p < .001, ηp2 = .56) and a significant reversed EC effect (F(1, 35) = 5.18, p = .03, ηp2 = .20) in the reversal condition. The size of the normal EC effect in the standard condition is almost three times as large as the size of the reversed EC effect in the reversal condition, as revealed by the partial η². An auxiliary test in which the sign of the EC effect is set positive in both standard and reversal conditions showed, however, that this size difference is not significant in this study (F(1, 35) = 2.35, p = .13, ηp2 = .06). MPT Model The frequency data from the dichotomous evaluation task were analyzed with the software HMMTree (Stahl and Klauer 2007; see appendix B for the frequency data of all experiments). The model fit the empirical data well (G²(1) = 0.19, p = .66), implying that the observed frequencies did not differ significantly from the frequencies predicted by the model. The estimate of the c-parameter was c = .28 (95% confidence interval (CI) [.21, .34]), thus reflecting 28% of the CS evaluations. The u-parameter’s estimate was u = .14, 95% CI [.06, .23], accounting for 10% of the pairings (i.e., the product of the converse probability of the c-parameter and the probability expressed by the u-parameter). Setting this parameter to zero led to a significant decrease in model fit (ΔG²(1) = 10.26, p < .001), indicating a significant contribution of the uncontrollable process to attitude formation. The parameter representing response tendencies was r = .44, 95% CI [.39, .50], indicating a slight tendency to respond “unpleasant” as r < .50; see figure 4 for the parameter estimates and their standard errors. FIGURE 4 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 1A NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 4 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 1A NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. Discussion The results clearly illustrate why the use of a MPT methodology is expedient to distinguish controllable from uncontrollable processes, as the results obtained with a classic measure of evaluation would be very hard to interpret unambiguously. On the one hand, we observe a reversed EC effect in the reversal condition, which could be interpreted as indicative of the controllable nature of EC. On the other hand, in the standard condition both the controllable and uncontrollable process lead to a regular EC effect, while in the reversal condition they have opposite effects on evaluations. Therefore, one could argue that the smaller size of the reversed EC effect in the reversal condition constitutes evidence for the uncontrollable nature of EC (although the difference in size was not significant in this study, it is significant in studies 2, 4, and 5). However, significant or not, evidence of this kind could not be regarded as conclusive, as alternative processes might lead to the same outcome (e.g., participants might be able to execute the reversals at will, but form less extreme evaluations for CSs when they perform the reversal). Crucially, the MPT model allows for dissociating and quantifying the contributions of controllable (c-parameter) and uncontrollable processes (u-parameter). Whereas a substantial subset of CSs acquired an evaluation in line with the instructed US valence as evidenced by the c-parameter, the experiment also yielded evidence for uncontrollable EC effects reflected by a significant u-parameter that captures evaluations in line with US valence irrespective of instructions. Thus, both types of processes concurrently contribute to evaluations. STUDY 1B: A WITHIN-PARTICIPANT TEST OF THE MPT MODEL In the next study, we implemented a within-participant operationalization of standard and reversal conditions. Obtaining similar results with a different operationalization of the method would enhance faith in the validity of the method. Specifically, it is possible that the c- and u-parameters could reflect standard and reversed responding at test rather than learning. A precondition for such response strategies would be that participants remember the US valence or acquired an EC effect (and respond accordingly in the standard condition, and give reversed responses in the reversal condition). As participants completed only one instruction condition in the between-participants design, it was theoretically possible to apply the instructions at test rather than during learning. By introducing a within-participant design that interspersed the CSs from both instruction conditions at test, we reduced this possibility, as participants would need to memorize additional information on the instructions pertaining to each CS to report the correct valence at test. If a reversal response set was responsible for the parameter structure obtained, the c-parameter should be smaller in the within- than in the between-participants design. An additional advantage of using a within-participant operationalization is that it provides the opportunity to estimate the c- and u-parameters for each participant individually. This, in turn, allows us to assess whether the parameter estimates predict participants’ ratings on an independent measure of evaluation. Observing that the parameter estimates co-vary as predicted with other measures of evaluation would enhance faith in the validity of the method by providing evidence of predictive validity. Finally, obtaining convergence of the parameter estimates across between- and within-participant operationalizations assuages concerns regarding violations of some assumptions underlying the MPT methodology (e.g., aggregation independence and parameter homogeneity, which are assumed when aggregating across participants; Curran and Hintzman 1995; Jacoby and Shrout 1997). Method Participants Twenty-three students of different majors at the University of Heidelberg took part in the study (Mage = 21.60, SDage = 2.80). We excluded three participants due to anomalous data in the dichotomous evaluation task, with one participant always reversing responses (even in the standard condition when no reversals were required), one responding “unpleasant” in 88% of the cases, and one responding “pleasant” in 92% of the cases. They were identified as outliers (Tukey 1977). The final sample contained 14 females and six males (as there are no between-participants manipulations, we considered this sample size sufficient to test the validity of the method; see our note on sample sizes and power preceding the General Discussion). Participants received course credit or €4.00 (approximately US$5.00). Design The study employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: preratings vs. postratings) × 2 (instruction condition: standard vs. reversal) design with repeated measures on all three factors. Sequence of instructions (standard-first vs. reversal-first) was counterbalanced between participants. Materials and Procedure The materials and procedure followed study 1a with one critical exception. Participants watched two conditioning blocks, one under standard and one under reversal instructions. Hence, the 24 CSs were split between the standard and reversal blocks, which contained 12 CSs each to be presented six times with a US. Dependent measures were administered only after both conditioning blocks were completed with CSs from both blocks commingled in random order in each measure. Results Evaluative Ratings Evaluative ratings were submitted to a 2 (US valence) × 2 (time of measurement) × 2 (instruction condition) × 2 (sequence of instructions) mixed ANOVA. The EC effect was significant (F(1, 18) = 5.17, p = .04, ηp2 = .22) and moderated by instruction condition (F(1, 18) = 20.74, p < .001, ηp2 = .54). Whereas there was no overall main effect of US valence (F(1, 18) = 0.39, p = .54, ηp2 = .02), the effect of US valence was dependent on instruction condition (F(1, 18) = 23.41, p < .001, ηp2 = .57). None of the other effects was significant (smallest p = .15). The EC effects were assessed separately for the two instruction conditions. The standard condition resulted in a significant EC effect (F(1, 19) = 41.77, p < .001, ηp2 = .68). The reversed EC effect in the reversal condition did not reach the conventional level of significance (F(1, 19) = 3.99, p = .06, ηp2 = .17). The partial η² reveal that the size of the EC effect in the standard condition is more than three times the size of the reversed EC effect in the reversal condition. An auxiliary test demonstrated this size difference was significant (F(1, 19) = 5.64, p = .03, ηp2 = .23). The evaluative ratings are displayed in figure 5. FIGURE 5 View largeDownload slide EVALUATIVE RATINGS IN STUDY 1B BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.—The error bars represent standard errors. FIGURE 5 View largeDownload slide EVALUATIVE RATINGS IN STUDY 1B BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.—The error bars represent standard errors. MPT Model The initial model containing one c-, one u-, and one r-parameter per sequence condition described the data well (G²(2) = 1.68, p = .43). The initial parameter estimates were c = .37, 95% CI [.25, .49], u = .15, 95% CI [.00, .35], and r = .46, 95% CI [.35, .58] for the standard-first condition. The parameter estimates for the reversal-first condition were c = .40, 95% CI [.29, .51], u = .19, 95% CI [.01, .37], and r = .52, 95% CI [.41, .63]. The u-parameter differed from zero in both the standard-first (ΔG²(1) = 2.24, p = .07) and reversal-first (ΔG²(1) = 4.01, p = .02) conditions. The c-parameters (ΔG²(1) = 0.15, p = .70), u-parameters (ΔG²(1) = 0.08, p = .78), and r-parameters (ΔG²(1) = 0.47, p = .49) can be set equal across sequence conditions without loss in model fit. The resulting parameter estimates were c = .39, 95% CI [.31, .47], u = .17, 95% CI [.04, .30] (which is significantly larger than zero, ΔG²(1) = 6.19, p < .01), and r = .49, 95% CI [.41, .57] (see figure 6). The final model fit the data well (G²(5) = 2.38, p = .79; frequency data are in appendix B). FIGURE 6 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 1B NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 6 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 1B NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. The MPT data can be modeled jointly for the between- (study 1a) and the within-participant version (study 1b). The initial model with one c-, one u-, and one r-parameter for each experiment showed a good fit (G²(2) = 1.87, p = .39). Setting the c-parameters equal across experiments impaired the model’s fit. In contrast to the reversal-at-response explanation, the c-parameter for the within-participant variant, cwithin = .39, 95% CI [.31, .47], was significantly larger than the one for the between-participants experiment, cbetween = .28, 95% CI [.21, .34], ΔG²(1) = 4.57, p = .03. The power analysis presented before the General Discussion indicates, however, that this difference between designs was relatively small. The u-parameters can be set equal without a loss in model fit, ΔG²(1) = 0.12, p = .73. The resulting u-parameter with an estimate of u = .15, 95% CI [.08, .23] was significantly larger than zero, ΔG²(1) = 16.70, p < .001. The estimates of the r-parameter also did not differ between conditions, ΔG²(1) = 0.98, p = .32, resulting in a combined estimate of r = .44, 95% CI [.42, .50]. The final model fit the data well, G²(4) = 2.97, p = .56. Predictive Value of Individual Parameter Estimates The next analysis investigates whether the parameter estimates obtained from the dichotomous evaluation task (in which participants rated CSs as pleasant or unpleasant) are predictive of CS evaluations on a separate, continuous measure. Standardized parameter values for the controllable and uncontrollable components were estimated per participant. We calculated EC effects from the continuous evaluative ratings by subtracting the mean evaluation of the negatively paired CSs from the mean evaluation of the CSs that have been paired positively. We specified a Hierarchical Linear Model (HLM) predicting the EC effects from the instruction condition, the c- and u-parameters, and their interaction with the instruction condition (the model is hierarchical, as the evaluative ratings for both instruction conditions are nested within participants). The main effect of instruction condition was significant, as the EC effect is positive in the standard condition, but negative in the reversal condition (B = –41.42, SE = 7.27, t(16) = –5.69, p < .001). The main effect of the c-parameter was not significant (B = –8.65, SE = 10.19, t(16) = –0.85, p = .41), but was moderated by instruction condition (B = –76.40, SE = 26.32, t(16) = –2.90, p = .01). As expected, the c-parameter shows a (marginally significant) positive relation to the EC effect in the standard condition (B = 29.55, SE = 16.64, t(16) = 1.78, p = .09) and a negative relation in the reversal condition (B = –46.85, SE = 16.65, t(16) = –2.81, p = .01). The u-parameter, on the other hand, is a universally positive predictor of the EC effect (B = 24.65, SE = 8.73, t(16) = 2.82, p = .01), irrespective of the instructions participants received (B = 23.12, SE = 22.55, t(16) = 1.03, p = .32). Discussion This study replicates study 1a in a within-participant design, increasing our confidence in the validity of the parameter estimates in three different ways. First, the fact that parameters were equal in size across sequence conditions alleviates potential concerns that switching between tasks would bias the parameters (Rogers and Monsell 1995). Second, the fact that the c-parameter was not impaired (but rather enhanced) in the within-participant variant suggests that participants executed the reversals during learning rather than at test. Third, the within-participant design enables estimating for every participant the contribution of controllable and uncontrollable processes from responses in the first, dichotomous evaluation task. Both parameter estimates predict the EC effect on a second, continuous measure of evaluation in the expected directions, demonstrating their predictive validity. Specifically, whereas the effect of the c-parameter depends on the instruction condition, the u-parameter has a universally positive influence on the EC effect, independent of instructions. The fact that the u-parameter successfully predicts evaluations over and above the c-parameter on a continuous direct measure of evaluation implies that the results obtained by direct measures are a combination of both controllable and uncontrollable processes. The (smaller) reversed EC effect observed on the direct measure in the reversal condition is thus composed of a mix of controllable, reversed EC effects and uncontrollable, standard EC effects. STUDY 2: MANIPULATION OF PROCESSING CAPACITY A key challenge of the current research is to establish the construct validity of the MPT parameters. We claim that the model’s c-parameter reflects the contribution of a controllable cognitive process, whereas the u-parameter reflects the contribution of an uncontrollable process. One can provide construct validity for MPT model parameters by demonstrating that the parameter values respond as predicted to experimental manipulations known to affect nonautomatic and automatic processes differently (Hütter et al. 2012). One primary variable of interest is working memory capacity. The controllability of a process should hinge on working memory capacity, as deliberative effort is required and negation operations are known to require extensive cognitive resources (Bargh 1994; Deutsch et al. 2009; Fazio 1990). Hence, our trust in the validity of the c-paramater as capturing a controllable cognitive process would be enhanced if its effect is reduced when cognitive capacity is impaired. In addition, manipulating cognitive capacity can provide evidence for the thesis that the u-parameter reflects an uncontrollable process that is independent of cognitive resources. This thesis would be supported if the u-parameter were equally large under conditions with full cognitive capacity as under conditions where cognitive capacity is constrained. Method Participants Fifty-nine students (49 female) of different majors at the University of Heidelberg took part in this experiment (Mage = 22.29, SDage = 4.50). Participants chose between course credit and €4.00 (approximately US$5.00). Design We employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: pre- vs. postratings) × 2 (instruction condition: standard vs. reversal) × 2 (cognitive load: low vs. high) mixed design with repeated measures on the first two factors. Materials and Procedure The materials were identical to those in studies 1a and 1b. We employed several procedural changes in order to establish a low- and a high-load condition. The conditioning phase was segmented into six blocks. In the high-load condition, a four-digit number was presented for six seconds before each block, which participants were asked to remember. In the low-load condition, no number was presented and participants saw the message “Please wait. To be continued shortly.” After each block, participants were asked to type in a four-digit number (the memorized one in the high-load condition; any number in the low-load condition). We decided to limit the number of digits to four mainly because the task of controlling evaluative learning is already demanding and previous research showed that high levels of cognitive and attentional demand can completely eliminate EC effects (Dedonder et al. 2010; Pleyers et al. 2009). Note that the goal of this manipulation was not to completely absorb processing resources but to demonstrate a relative dependence on processing resources of one process that is not present in the other. The sequence of CSs within each block was randomized. Results Evaluative Ratings Evaluative ratings, illustrated in figure 7, were submitted to a 2 (US valence) × 2 (time of evaluative rating) × 2 (instruction condition) × 2 (cognitive load) mixed ANOVA. The overall EC effect (i.e., the US valence × time of measurement interaction) was significant (F(1, 55) = 4.54, p = .04, ηp2 =.08), involving also a significant main effect of US valence (F(1, 55) = 6.62, p = .01, ηp2 = .11). The EC effect was significantly moderated by instruction condition (F(1, 55) = 12.82, p < .001, ηp2 = .19) as was the main effect of US valence (F(1, 55) = 10.62, p < .01, ηp2 = .16). Whereas the EC effect was not moderated by cognitive load (F(1, 55) = 0.00, p = 1, ηp2 = 0.00), its interaction with instruction condition was significantly moderated by load (F(1, 55) = 5.73, p = .02, ηp2 = 0.02). There was a marginally significant main effect of instruction condition (F(1, 55) = 3.48, p = .07, ηp2 = .09). All other effects were nonsignificant (largest F = 2.48, smallest p = .12). FIGURE 7 View largeDownload slide EVALUATIVE RATINGS IN STUDY 2 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND COGNITIVE LOAD CONDITION NOTE.— The error bars represent standard errors. FIGURE 7 View largeDownload slide EVALUATIVE RATINGS IN STUDY 2 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND COGNITIVE LOAD CONDITION NOTE.— The error bars represent standard errors. To decompose the four-way interaction, we calculated separate 2 (US valence) × 2 (time of evaluative rating) ANOVAs for each cell of the between-participant design (load × instruction condition). Without cognitive load, there was a regular EC effect in the standard condition (F(1, 14) = 39.80, p < .001, ηp2 = .74), while the reversed EC effect in the reversal condition was not significant (F(1, 13) = 2.39, p = .15, ηp2 = .16). Under cognitive load, the standard condition also resulted in a reliable EC effect (F(1, 14) = 5.76, p = .03, ηp2 = .29), while the pairings had no effect in the reversal condition (F(1, 14) = 0.14, p = .72, ηp2 = .01). In the low-load condition, the size of the EC effect in the standard condition was more than twice the size of the reversed EC effect in the reversal condition, as revealed by the partial η². However, this difference failed to reach the conventional significance level in this study (F(1, 27) = 2.41, p = .13, ηp2 = .08). MPT Model An initial model was fit to the frequency data (in appendix B) containing one c-, one u-, and one r-parameter per load condition. Initial model fit was good (G²(2) = 2.62, p = .27). Estimates of the parameters in the low-load condition were c = .28, 95% CI [.21, .35], u = .13, 95% CI [.03, .23], and r = .48, 95% CI [.42, .54]. In the high-load condition, they were c = .07, 95% CI [.00, .14], u = .07, 95% CI [.00, .15], and r = .50, 95% CI [.45, .54]. See figure 8 for the parameter estimates and their standard errors. The individual u-parameters differed significantly from zero in both the low-load (ΔG² (1) = 6.74, p < .01) and the high-load conditions (ΔG² (1) = 3.22, p = .04). While the u-parameters (ΔG²(1) = 0.82, p = .37) and the r-parameters (ΔG²(1) = 0.23, p = .63) did not differ between the cognitive load conditions, the estimate for the c-parameter was significantly larger in the low-load condition than in the high-load condition (ΔG²(1) = 15.62, p < .001). The resulting model containing the four parameters clow = .28, 95% CI [.21, .35], chigh = .07, 95% CI [.00, .15], u = .09, 95% CI [.03, .16], and r = .49, 95% CI [.46, .52], fit the data well (G²(4) = 3.67, p = .45). The global u-parameter was significantly larger than zero (ΔG² (1) = 9.07, p < .01). FIGURE 8 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 2 NOTE.— c denotes the controllable parameter with separate estimates for the high-load and the low-load conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 8 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 2 NOTE.— c denotes the controllable parameter with separate estimates for the high-load and the low-load conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. Discussion A cognitive load manipulation was introduced to establish construct validity of the c- and u-parameters as capturing an effortful, controllable process versus an efficient, uncontrollable process, respectively. As expected, the estimate of the c-parameter was significantly reduced in the high-load condition, while estimates of the u-parameter were equal in size across load conditions. Hence, the process represented by the u-parameter is characterized by a core feature of automaticity: it is unaffected by variations in cognitive capacity (Bargh 1994). This finding supports the notion that the u-parameter does not merely reflect a deliberate application of US valence irrespective of task instructions. In this study, the standard and reversal conditions were manipulated between participants, precluding a regression analysis of the direct evaluations on the MPT parameter estimates (as in study 1a, only aggregated c- and u-parameter estimates are obtained). Nevertheless, it is illustrative to see how the pattern of EC effects observed on the direct evaluations is consistent with the obtained MPT parameter values on another measure. Remember that in the standard condition, controllable and uncontrollable processes act in concert to produce a regular EC effect. In the reversal condition, both processes act in opposition. While the u-parameter does not differ between the low-load and high-load conditions, the c-parameter is significantly reduced under high load. In the low-load condition, we thus observe a regular EC effect in the standard condition produced by both uncontrollable and controllable processes. As the c-parameter is larger than the u-parameter, controllable processes successfully counteract the influence of uncontrollable ones in the reversal condition, producing a significant, descriptively smaller, reversed EC effect. In the high-load condition, the c-parameter is significantly reduced and thus adds less to the EC effect in the standard condition and less powerfully contradicts uncontrollable processes in the reversal condition. Hence, the regular EC effect in the standard condition is considerably smaller and the opposition of processes in the reversal condition, which are about equally strong under high load, results in a null effect. STUDY 3: MANIPULATION OF MOTIVATION TO CONTROL Study 3 was designed to provide additional evidence for the construct validity of the parameters by manipulating a second crucial variable known to distinguish between controllable and uncontrollable processes—namely, motivation. Consumers’ level of motivation to execute control should affect only controllable processes, leaving uncontrollable processes unaffected (Bargh 1994; Fazio 1990). To investigate the influence of motivation on the controllability of EC, we presented the conditioning phase as a performance task, and manipulated the presence of financial incentives for forming “correct” impressions. Financial incentives should maximize the likelihood that participants try their best to execute the reversals. If truly reflective of a controllable process, the c-parameter should be greater in the financial incentives condition as compared to a control condition. Moreover, in this setup uncontrollable attitude acquisition has negative, undesirable consequences for participants’ financial compensation. To the extent that a significant u-parameter is obtained even in this condition, its interpretation as reflecting uncontrollable evaluative learning will be strengthened. This experiment also serves the secondary goal of strengthening the predictive validity evidence that was obtained in study 1b. In that study, we tested whether the individual c- and u-parameter estimates predict participants’ responses on a separate, continuous measure of evaluation. Whereas the results confirmed our expectations, the sample size in study 1b was relatively small. Therefore, again we implemented a within-participant design and increased the number of participants to estimate the individual parameters with greater accuracy. Method Participants Seventy-four students of different majors at a German university took part in the experiment. Five participants had not understood the instructions (i.e., they executed no reversals in the reversal condition) and were excluded from analysis. This left 69 participants (41 females, 28 males) for analysis (Mage = 24.22 years, SDage = 6.98). Participants received course credit or financial compensation of €3.00. In the financial incentives condition, participants received an additional performance-based reward (see below for details). Participants in the control condition were also paid €2.70 extra (corresponding to 75% correct classifications), but were informed about this only after the experiment. Design The experiment employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: preratings vs. postratings) × 2 (instruction condition: standard vs. reversal) × 2 (financial incentives: control vs. incentives) mixed design with repeated measures on the first three factors. Materials and Procedure The implementation of a performance task required instructions that were more explicit about the usually observed effects of the pairings and participants’ task. We thus implemented instructions that were similar to the standard and reversal instructions used in study 2 (see appendix A for the exact wording of the instructions). Additionally, participants in the financial incentives condition received instructions on a performance-based reward. The standard block was always presented first and followed by the reversal block. We decided on this fixed sequence to reduce error variance and because the sequence factor did not influence the results in study 1b. Results Evaluative Ratings Evaluative ratings, displayed in figure 9, were submitted to a 2 (US valence) × 2 (time of evaluative rating) × 2 (instruction condition) mixed ANOVA. The overall EC effect was significant (F(1, 67) = 5.31, p = .02, ηp2 = .07), as was the main effect of US valence (F(1, 67) = 4.23, p = .04, ηp2 = .06). There also was a significant main effect of time of measurement (F(1, 67) = 4.44, p = .04, ηp2 = .04). The EC effect was moderated by instruction condition (F(1, 67) = 43.45, p < .001, ηp2 = .39), as was the main effect of US valence (F(1, 67) = 32.03, p < .001, ηp2 = .32). None of the other effects was significant (smallest p = .16). FIGURE 9 View largeDownload slide EVALUATIVE RATINGS IN STUDY 3 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND FINANCIAL INCENTIVES CONDITION NOTE.— The error bars represent standard errors. FIGURE 9 View largeDownload slide EVALUATIVE RATINGS IN STUDY 3 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND FINANCIAL INCENTIVES CONDITION NOTE.— The error bars represent standard errors. Separate analyses of the EC effects in the standard and reversal conditions demonstrate that the regular EC effect in the standard condition was significant (F(1, 67) = 52.91, p < .001, ηp2 = .44) and not moderated by incentives condition (F(1, 67) = 1.13, p = .29, ηp2 = .02). While the reversed EC effect in the reversal condition was much smaller than the standard effect, it was also significant (F(1, 67) = 13.69, p < .01, ηp2 = .17), irrespective of incentives condition (F(1, 67) = 1.05, p = .31, ηp2 = .02). As reflected by the partial η², the size of the EC effect in the standard condition was more than twice as large as the size of the (reversed) EC effect in the reversal condition. This size difference was statistically significant (F(1, 68) = 5.45, p = .02, ηp2 = .07). MPT Model The initial model containing one c-, one u-, and one r-parameter per incentives condition demonstrated a good fit to the data (G²(2) = 0.55, p = .76) and is presented in figure 10. The initial parameter estimates in the control condition were c = .33, 95% CI [.26, .39], u = .10, 95% CI [.00, .19] (significantly different from zero, ΔG²(1) = 4.24, p = .02), and r = .49, 95% CI [.44, .55]. The parameter estimates for the financial incentives condition were c = .61, 95% CI [.55, .66], u = .15, 95% CI [.00, .29] (significantly different from zero, ΔG²(1) = 4.03, p = .02), and r = .50, 95% CI [.42, .59]. While the u-parameters did not differ between incentives conditions (ΔG²(1) = 0.33, p = .57), the c-parameters were different (ΔG²(1) = 42.61, p < .001). The r-parameters also did not differ (ΔG²(1) = 0.03, p = .86). FIGURE 10 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 3 NOTE.— c denotes the controllable parameter with separate estimates for the control and incentives conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars indicate the standard error of the individual parameter estimates. FIGURE 10 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 3 NOTE.— c denotes the controllable parameter with separate estimates for the control and incentives conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars indicate the standard error of the individual parameter estimates. Predictive Value of Individual Parameter Estimates As in study 1b, we conducted an HLM predicting the EC effect as a function of participants’ individual c- and u-parameter estimates (standardized), the instruction condition, the financial incentives condition, and the parameters’ interactions with the instruction and incentives conditions. For 10 participants, the u-parameter was unidentified (this occurs when some category frequencies are too close to zero). The main effect of instruction condition was again significant (B = –35.94, SE = 4.92, t(54) = –7.31, p < .001), reflecting the fact that the EC effect is positive in the standard condition, but negative in the reversal condition. The main effect of the c-parameter was marginally significant (B = –4.60, SE = 2.37, t(53) = –1.94, p = .058), but this effect was moderated by instructions (B = –23.57, SE = 5.50, t(54) = –4.28, p < .001). As expected, the c-parameter showed a positive relation to the EC effect in the standard condition (B = 7.19, SE = 3.63, t(53) = 1.98, p = .05) and a negative relation in the reversal condition (B = –16.38, SE = 3.63, t(53) = –4.51, p < .001). The u-parameter was a universally positive predictor of the EC effect (B = 8.64, SE = 2.04, t(53) = 4.24, p < .001), as its effect was not moderated by instruction condition (B = 1.54, SE = 4.80, t(54) = 0.32, p = .75). Finally, the financial incentives manipulation did not have a direct effect on EC, nor did it qualify the (main or interactive) effects of the c- and u-parameters on EC (smallest p = .13). Discussion This study provides several important pieces of evidence for the existence of uncontrollable EC effects. First, the study strengthens the predictive validity of the MPT parameters. As in study 1b, the HLM analysis indicates that the c- and u- parameters correlate as predicted with an independent measure of the EC effect. Second, the study assuages a potential concern that participants in previous studies might not have been sufficiently motivated to comply with task instructions. In this study’s financial incentives condition, it was financially costly for participants to acquire evaluations in an uncontrollable manner. Yet, even in this condition, EC continued to exert an uncontrollable influence on the CS evaluations, as evidenced by the significant u-parameter. Third, and most importantly, the study enhances our faith in the construct validity of the u- parameter as capturing an uncontrollable process. As demonstrated by its effect on the c-parameter, the financial incentives manipulation did increase participants’ likelihood to exert control. At the same time, however, financial incentives did not influence the u-parameter. This implies that in those cases where control fails, there remains a constant probability of an uncontrollable effect of EC, unaffected by the motivation to exert control. This process’s resistance to effortful control supports the notion that this parameter does not represent the deliberate application of US valence regardless of task instructions and is a hallmark property of its automatic nature. In this study, we were highly explicit about the performance criterion necessary to achieve additional rewards. While this study supports the existence of an uncontrollable learning mechanism, the study’s emphasis on performance imposes a limitation on the interpretation of the c-parameter. It is possible that in some cases, the c-parameter does not capture genuine evaluative learning, but rather a demand effect of participants reporting the correct CS valence as requested by the instructions. Demand artifacts are a notable source of concern in many experimental settings, and particularly so in settings where the researcher’s hypotheses and expectations are transparent for participants (Page 1969). Note, however, that one desirable and crucial advantage of the method we developed is that while demand effects could inflate the size of the c-parameter, they go against finding a significant u-parameter, which is derived from participants’ inability to conform to instructions or task demands. STUDY 4: EXTERNAL VALIDITY OF PARADIGM The purpose of the next study was to increase the external validity of our paradigm in two different ways. All previous studies used the same stimulus materials, working with naturalistic faces as CSs. In the next study, we employed consumer brands as CSs to investigate whether our results hold in the consumer context, as we would claim. Second, our previous studies were primarily concerned with the internal validity of our research paradigm. To demonstrate the internal validity of this research, it is important that participants are instructed to exert control to the best of their ability. This is the reason that in all previous studies, we explicitly encouraged participants to either apply or reverse the valence of the US in forming their attitude toward the CS. As is often the case, however, maximizing internal validity comes at the cost of reducing the external validity—that is, the generalizability of the findings to other paradigms and settings. Specifically, the instructions in EC research more generally often conceal the purpose of the CS-US pairings (Olson and Fazio 2001; Sweldens et al. 2010). The question then arises whether our method is also informative about such contexts, and what would be the estimated contribution of controllable and uncontrollable processes when participants are not instructed to exert control. Analogous to the previous studies, we included a standard and reversal condition with highly explicit instructions to apply or reverse the influence of US valence, respectively. Additionally, we included a “covert condition” with instructions that conceal the purpose of the pairings, thereby reducing the probability that participants exert control over the attitude acquisition. Method Participants We collected data of 173 international students at Erasmus University Rotterdam. We excluded two participants (1.16 %) who gave the same response in the MPT task in 88% of the cases and above. Of these participants, 94 were women and 97 were men. They were between 17 and 24 years old (Mage = 19.94, SDage = 1.25). Design The experiment implemented a 2 (US valence: positive vs. negative) × 2 (time of measurement: pre- vs. post-pairings) × 3 (instructions: standard vs. reversal vs. covert) with repeated measures on the first two factors. Materials and Procedure The CS repertoire consisted of 38 bottled water brand logos that were specifically designed for this experiment and thus unknown to participants. For each participant, the software selected the 24 CSs that received the most neutral preratings. The instructions were adapted to the materials and were very explicit about the goal of the task (exerting control) in the standard and reversal conditions. In the covert condition, we employed instructions that were used by Sweldens and colleagues (2010). These instructions asked participants to familiarize themselves with the brands and concealed the purpose of the pairings by stating that the US images were presented to make the presentation of the CS brands more interesting (the exact wording can be found in appendix A). Results Evaluative Ratings The evaluative ratings revealed a large EC effect (F(1, 168) = 65.17, p < .001, ηp2 = .28) that was moderated by condition (F(2, 168) = 7.93, p < .001, ηp2 = .09) as apparent in figure 11. The EC effect was large in both the covert (F(1, 56) = 36.08, p < .001, ηp2 = .39) and the standard conditions (F(1, 54) = 31.31, p < .001, ηp2 = .37) and was reduced to a small but regular EC effect in the reversal condition (F(1, 58) = 4.14, p < .05, ηp2 = .07). The absolute sizes of the EC effects did not differ between the standard and covert conditions (F(1, 110) = 2.05, p = .16, ηp2 = .02), while effect sizes differed between the standard and reversal conditions (F(1, 112) = 13.57, p < .001, ηp2 = .11), as well as the covert and the reversal conditions (F(1, 114) = 8.24, p < .01, ηp2 = .07). FIGURE 11 View largeDownload slide EVALUATIVE RATINGS IN STUDY 4 BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. FIGURE 11 View largeDownload slide EVALUATIVE RATINGS IN STUDY 4 BY US VALENCE, TIME OF MEASUREMENT, AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. The pattern of means across the three conditions suggests that in this experiment, the instructions mostly affected the positively paired CSs. Indeed, the moderation of the EC effect by instruction condition was significant only for positively paired CSs (F(2, 168) = 8.20, p < .001, ηp2 = .09), but not for negatively paired CSs (F(2, 168) = 1.99, p = .14, ηp2 = .02). This suggests that control was not successfully applied when US images were negative. The MPT model will test for this assumption by comparing the c-parameters between the positive and negative US valence conditions. MPT Model We first analyzed the MPT model for the standard and reversal conditions only. The estimates of the c-parameter differed significantly between positive and negative US valences, ΔG²(1) = 16.05, p < .001. These parameters thus must be estimated separately. The c-parameter in the negative pairings condition amounted to c- = .05, 95% CI [.00, .10], and in the positive pairings condition to c+ = .20, 95% CI [.15, .25]. The estimate of the u-parameter was u = .18, 95% CI [.14, .22], and differed significantly from zero, ΔG²(1) = 70.02, p < .001. The r-parameter amounted to r = .53, 95% CI [.51, .56]. This slight bias toward “pleasant” was significant, ΔG²(1) = 5.84, p = .02. With four parameters this model is saturated and thus does not allow for an assessment of fit, but the fact that the G² statistic was not larger than zero indicates that this model’s equality restrictions (e.g., that we expected more responses to oppose US valence in the reversal condition) were met, G²(0) = 0.00. The parameter estimates are depicted in figure 12. FIGURE 12 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 4 NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 12 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 4 NOTE.— c denotes the controllable parameter; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. In the next step, we calculated the MPT model for all three experimental conditions, including the covert condition with the more typical EC instructions. The u-parameter amounted to u = .16 [.12, .20], and was significantly larger than zero, ΔG²(1) = 68.39, p < .001. Furthermore, it did not differ significantly from its estimate when we used the data of the standard and reversal conditions only, ΔG²(1) = 1.50, p = .22. The estimate of the u-parameter was thus not altered by the inclusion of the 1,368 observations from the covert condition. Again we estimated the c-parameters of the standard and reversal conditions separately for negative, c- = .04, 95% CI [.00, .10], ΔG²(1) = 2.70, p = .05, and positive pairings, c+ = .19, 95% CI [.14, .24], ΔG²(1) = 54.03, p < .001, as they differed strongly from another, ΔG²(1) = 15.98, p < .001. The estimate of the c-parameter in the covert condition amounted to ccovert = .00, 95% CI [.00, .08], and did not differ from zero, ΔG²(1) = 0.00, p = 1. The r-parameter amounted to r = .53, 95% CI [.51, .55], and thereby indicated a slight tendency to respond “pleasant” in this task, ΔG²(1) = 9.46, p < .01. With the addition of the covert condition, we gained degrees of freedom allowing for an assessment of model fit. The final model across all conditions described the data well, G²(1) = 2.66, p = .10. Discussion The present study allows several comparisons between explicit control instructions and instructions that covered the true purpose of the pairings. When calculating a joint model for all three conditions, we restricted the u-parameter to be equal across conditions. Nevertheless, its value could have changed due to the inclusion of 1,368 additional observations. The fact that its estimate was not affected is consistent with its interpretation as reflecting the uncontrollable influence of US pairings. Moreover, the c-parameter that could vary freely was strongly affected by the presence versus absence of explicit control instructions, as it was reduced to insignificance in the covert condition, where participants were not instructed to exert control. Hence, the present results suggest that uncontrollable learning may be relatively stable across different EC paradigms, while the degree of controllable learning may vary considerably. The fact that the MPT parameters respond as predicted to the presence versus absence of control instructions is further evidence for their construct validity. Furthermore, the study provides evidence for the relevance of our approach to the wider EC literature. The EC effect in the standard and covert conditions was approximately equal in size, assuaging concerns that the EC effect in our studies might be completely dependent on the presence of explicit control instructions. In fact, our data even indicate that controllable processing might play a negligible role in typical EC paradigms, as evidenced by the fact that ccovert did not differ from zero. One finding we did not anticipate was the difference in control exertion between positive and negative US pairings, which was not present in any other study. The fact that participants exerted less control over negative US pairings might be consistent with the so-called “negativity bias,” in which negative information has a stronger effect on evaluations than comparable positive information (Ito et al. 1998). We do not know, however, why this bias manifested itself solely in this study and could only speculate about the various slight differences in sample characteristics and in the combination of conditioned stimuli (consumer brands) and instructions unique to this study (see appendix A for the differences). STUDY 5: PREDICTING PRODUCT CHOICE AND CONSUMPTION This study’s main goal was to investigate whether the uncontrollable EC effect on attitudes documented in the previous studies influences actual product choices and consumption. Apart from its primary substantive importance, finding evidence for a relation between the MPT model parameters and actual product choices would establish a crucial part of the predictive validity of the MPT modeling approach. To study the effects on product choices, we again used the initially neutral logos of bottled water brands as our CS repertory, which would be paired with positive versus negative pictures during the conditioning phase. We then assessed the choices of participants when they were given the opportunity to taste these waters. Moreover, we also assessed the amount they consumed as an unobtrusive measure of approach behavior. Method Participants Eighty-one students of different majors at the University of Tübingen took part in this experiment (53 female; Mage = 23.14, SDage = 4.87). The MPT modeling was based on 1,944 observations. Design The experiment employed a 2 (US valence: positive vs. negative) × 2 (time of measurement: preratings vs. postratings) × 2 (instruction condition: standard vs. reversal) design with repeated measures on all three factors. Sequence of instructions (standard-first vs. reversal-first) was counterbalanced between participants. Materials and Procedure With the exception that the covert condition was omitted from this study, materials and procedure were identical to the previous study. Going beyond the previous study, participants were invited to refresh themselves with the waters of the brands presented during the study in an adjacent room after the computer-based part was finished. Plastic cups that were labeled with the CS logos were lined up on a table. Each cup was filled with 130 milliliters of tap water. Participants were told that they could drink as little or as much from the cups as they liked. An experimenter who was blind to the assignment of CSs to conditions recorded the cups selected by each participant. After the participant left, the amount consumed was determined with a calibrated digital weighing scale. Results Evaluative Ratings The EC effect was significant, F(1, 80) = 28.77, p < .001, ηp2 = .27, but qualified by instruction condition, F(1, 80) = 88.78, p < .001, ηp2 = .53. Both the regular EC effect in the standard condition, F(1, 80) = 119.54, p < .001, ηp2 = .60, and the reversed EC effect in the reversal condition, F(1, 80) = 6.87, p <.05, ηp2 = .08, were significant. The difference in the absolute sizes of the EC effect was significant, F(1, 80) = 28.77, p < .001, ηp2 = .27 (figure 13). FIGURE 13 View largeDownload slide EVALUATIVE RATINGS IN STUDY 5 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND FINANCIAL INCENTIVES CONDITION. NOTE.— The error bars represent standard errors. FIGURE 13 View largeDownload slide EVALUATIVE RATINGS IN STUDY 5 BY US VALENCE, TIME OF MEASUREMENT, INSTRUCTION CONDITION, AND FINANCIAL INCENTIVES CONDITION. NOTE.— The error bars represent standard errors. Consumption Behavior We analyzed the number of cups chosen and the total amount of water consumed in two separate 2 (US valence) × 2 (instruction condition) repeated-measures ANOVAs as displayed in figure 14. Participants’ choices of cups in the tasting procedure demonstrated a main effect of US valence (F(1, 80) = 24.14, p < .001, ηp2 = .23), a main effect of instruction condition (F(1, 80) = 5.40, p = .02, ηp2 = .06), and their interaction (F(1, 80) = 72.89, p < .001, ηp2 = .48). Planned contrasts revealed that participants selected more cups of positively than negatively paired brands in the standard condition (F(1, 80) = 94.26, p < .001, ηp2 = .54), which was reversed in the reversal condition (F(1, 80) = 11.68, p < .001, ηp2 = .13). Results from the consumption measure were equivalent. The total amount of water consumed was a function of US valence × instruction condition (F(1, 80) = 53.06, p < .001, ηp2 = .40), indicating a regular EC effect in the standard condition (F(1, 80) = 61.41, p < .001, ηp2 = .43) and a reversed EC effect in the reversal condition (F(1, 80) = 7.74, p < .01, ηp2 = .09). FIGURE 14 View largeDownload slide BEHAVIORAL MEASURES TAKEN IN STUDY 5 BY US VALENCE AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. FIGURE 14 View largeDownload slide BEHAVIORAL MEASURES TAKEN IN STUDY 5 BY US VALENCE AND INSTRUCTION CONDITION NOTE.— The error bars represent standard errors. MPT Model The parameter estimates and standard errors are displayed in figure 15. They amounted to c = .48, 95% CI [.44, .51], u = .38, 95% CI [.31, .45], and r = .47, 95% CI [.41, .52]. The u-parameter was larger than zero, ΔG²(1) = 101.35, p < .001. The model fit the data well (G²(1) = 1.62, p = .20; frequency data in appendix B). FIGURE 15 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 5 NOTE.— c denotes the controllable parameter with separate estimates for the high-load and the low-load conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. FIGURE 15 View largeDownload slide PARAMETER ESTIMATES OF THE MPT MODEL FOR STUDY 5 NOTE.— c denotes the controllable parameter with separate estimates for the high-load and the low-load conditions; u denotes the uncontrollable parameter; r indicates the response tendency toward “pleasant.” The error bars denote the standard errors of the parameter estimates. Predictive Value of Individual Parameter Estimates We conducted three HLMs to assess the (standardized) c- and u-parameters’ impact on the attitudinal EC effect (i.e., the difference in attitudes between positively and negatively conditioned brands), the brand choice EC effect (i.e., the difference in cups chosen of these brands), and the consumption EC effect (i.e., the difference in water consumption between these brands). Six participants had unidentified u- and r-parameters (this occurs when some category frequencies are close to zero). The analysis on evaluative ratings revealed a significant main effect of instruction condition, as the EC effect is positive in the standard condition, but reversed in the reversal condition (B = –64.46, SE = 5.25, t(71) = –12.29, p < .001). We also observed a significant main effect of the c-parameter (B = –6.93, SE = 2.73, t(71) = –2.54, p = .01), which was moderated by instruction condition (B = –39.29, SE = 5.57, t(71) = –7.06, p < .001). While the c-parameter predicted a regular EC effect in the standard condition (B = 12.71, SE = 3.90, t(71) = 3.26, p < .01), it generated a reversed EC effect in the reversal condition (B = –26.58, SE = 3.90, t(71) = –6.82, p < .001). Conversely, the u-parameter had a universally positive influence on the EC effect (B = 17.72, SE = 2.59, t(71) = 6.84, p < .001), which was not moderated by reversal instructions (B = –3.14, SE = 5.29, t(71) = –0.59, p = .55). The analysis on brand choice revealed a main effect of instruction condition (B = –2.80, SE = 0.29, t(71) = –9.56, p < .001) and a main effect of the c-parameter (B = –0.31, SE = 0.12, t(71) = –2.59, p = .01). Importantly, the latter was qualified by instruction condition (B = –1.60, SE = 0.31, t(71) = –5.13, p < .001). As expected, the c-parameter had a positive influence on brand choice in the standard condition (B = 0.49, SE = 0.20, t(71) = 2.48, p = .02) and a negative influence in the reversal condition (B = –1.11, SE = 0.20, t(71) = –5.64, p < .001). Again, the u-parameter had a universally positive influence (B = 0.62, SE = 0.11, t(71) = 5.41, p < .001), unaffected by instructions (B = –0.09, SE = 0.30, t(71) = –0.31, p = .76). The analysis of water consumption (in milliliters) mirrored the analysis on brand choice. Again, the u-parameter had a universally positive influence (B = 12.55, SE = 5.62, t(71) = 2.23, p = .03), unaffected by instructions (B = 3.21, SE = 10.58, t(71) = 0.30, p = .76). The c-parameter’s effect was moderated by instructions (B = –43.54, SE = 11.13, t(71) = –3.91, p < .001), with a positive weight in the standard condition (B = 13.35, SE = 8.12, t(71) = 1.64, p = .10) and a negative weight in the reversal condition (B = –30.19, SE = 8.12, t(71) = –3.72, p < .001). Discussion Study 5 goes beyond the demonstration of uncontrollable evaluative learning and establishes the predictive validity of the MPT parameters for product choice and consumption. The present results clearly demonstrate the uncontrollable impact of visual affective stimulus pairings on consumption behavior that comprised both the choice of drinks and their consumption. Participants often selected and consumed brands that were presented with positive contents, and did not consume brands that were presented with negative images, even though they were instructed to reverse the impact of the pairings on their evaluations. Hence, even when participants knew about the deceptive nature of advertising, neither their evaluations nor their consumption behavior were immune to its content. SAMPLE SIZES AND POWER IN STUDIES 1–5 Sample sizes were based on previous research using MPT modeling (Hütter and Sweldens 2013; Hütter et al. 2012). In all studies, each participant provided 24 observations (i.e., responses towards 24 CSs) as a basis for estimating the MPT model. In the first experiments testing merely the structure of the model, we thus aimed for up to 40 (study 1a) and 20 participants (study 1b), respectively. These sampling strategies resulted in 480 observations even in the study with the smallest sample size (study 1b; see also appendix B). We increased sample size considerably in all studies that introduced between-participants factors and those that assessed correlations between measures. As apparent from table 1, power analyses support our substantive conclusions in all experiments. Even in study 1b, the power to detect a u-parameter amounted to .70 and the power of detecting an influence of uncontrollable learning exceeds the conventional standard of .80 in all the other studies. Furthermore, the lack of an effect of attentional load (study 2) and financial incentives (study 3) on the u-parameter cannot be explained by a weak or unsuccessful manipulation, as they clearly affected the c-parameter (and evaluative ratings in study 2). The effect sizes of the manipulations’ effects on the c-parameters were in all cases considerably larger than the effects on the u-parameters, which were negligible. Table 1 Power Analyses for Central Hypothesis Tests Significance of u-parameter (test against zero) Study Parameter Cohen’s w N Power (1 – β) 1a u 0.11 888 .90 1b u 0.11 480 .70 2 u 0.07 1656 .80 3 u 0.08 1416 .85 4 u 0.17 4104 1 5 u 0.23 1644 1 Parameter comparisons across designs and experimental conditions 1a/1b: within versus between c 0.06 1368 .56 u 0.01 .06 2: low versus high load c 0.16 1656 1 u 0.01 .09 3: control versus incentives c 0.11 1416 .98 u 0.02 .15 Significance of u-parameter (test against zero) Study Parameter Cohen’s w N Power (1 – β) 1a u 0.11 888 .90 1b u 0.11 480 .70 2 u 0.07 1656 .80 3 u 0.08 1416 .85 4 u 0.17 4104 1 5 u 0.23 1644 1 Parameter comparisons across designs and experimental conditions 1a/1b: within versus between c 0.06 1368 .56 u 0.01 .06 2: low versus high load c 0.16 1656 1 u 0.01 .09 3: control versus incentives c 0.11 1416 .98 u 0.02 .15 NOTE.— N = number of observations. The rule of thumb for effect size w is that w = .1 is “small,” w = .3 is “medium,” and w = .5 is “large” (Cohen 1988). Power analyses were conducted with multiTree (Moshagen 2010). Table 1 Power Analyses for Central Hypothesis Tests Significance of u-parameter (test against zero) Study Parameter Cohen’s w N Power (1 – β) 1a u 0.11 888 .90 1b u 0.11 480 .70 2 u 0.07 1656 .80 3 u 0.08 1416 .85 4 u 0.17 4104 1 5 u 0.23 1644 1 Parameter comparisons across designs and experimental conditions 1a/1b: within versus between c 0.06 1368 .56 u 0.01 .06 2: low versus high load c 0.16 1656 1 u 0.01 .09 3: control versus incentives c 0.11 1416 .98 u 0.02 .15 Significance of u-parameter (test against zero) Study Parameter Cohen’s w N Power (1 – β) 1a u 0.11 888 .90 1b u 0.11 480 .70 2 u 0.07 1656 .80 3 u 0.08 1416 .85 4 u 0.17 4104 1 5 u 0.23 1644 1 Parameter comparisons across designs and experimental conditions 1a/1b: within versus between c 0.06 1368 .56 u 0.01 .06 2: low versus high load c 0.16 1656 1 u 0.01 .09 3: control versus incentives c 0.11 1416 .98 u 0.02 .15 NOTE.— N = number of observations. The rule of thumb for effect size w is that w = .1 is “small,” w = .3 is “medium,” and w = .5 is “large” (Cohen 1988). Power analyses were conducted with multiTree (Moshagen 2010). GENERAL DISCUSSION Consumers are constantly exposed to advertisements featuring visual imagery with affective content, which comes as no surprise, as it has long been established that pairing brands with appetitive stimuli is an effective method to increase brand liking (Bierley et al. 1985; Stuart et al. 1987). Very little research exists that investigates the controllability of attitude acquisition via such pairings. In the current research, we develop and validate a method to separate and quantify controllable and uncontrollable attitude acquisition in evaluative conditioning. We report six experiments, which demonstrate that simple presentations of visual affective stimuli can have uncontrollable effects on consumers’ attitudes and behavior. We discuss the theoretical, methodological, and substantive implications of our findings, as well as their limitations. Methodological Contributions Over the past two decades, consumer research has had a strong focus on—and made great progress in—studying implicit processes in memory, affect, and persuasion (Johar, Maheswaran, and Peracchio 2006). Standards for the investigation of implicit processes have been discussed intensely recently (Sweldens et al. 2017; Williams and Poehlman 2017). Even though the processing tree framework has many advantages over conventional approaches, applications thereof have been surprisingly uncommon in consumer research, with a few notable exceptions (Dalton and Huang 2014; Fitzsimons and Williams 2000; Pham and Johar 1997; Shapiro 1999). Nevertheless, further progress can still be made and a more widespread adoption of the processing tree framework, accompanied with more up-to-date methods of analysis, has the potential to improve our knowledge in this fascinating subject area significantly. Challenges of Previous Approaches to Uncontrollability Only very recently, research has started to explore whether pairings with visual affective stimuli can have uncontrollable effects on attitudes. Two articles by Gawronski and colleagues (2014, 2015) demonstrated that EC procedures have uncontrollable effects on indirect attitude measures, while effects on direct measures could still be controlled. As we outlined in more detail in the introduction, we believe more research on this topic is necessary for both substantive and methodological reasons. Substantively, EC effects are fully uncontrollable when they fail to be corrected at both the encoding stage and the validation stage—that is, at retrieval and application of these evaluations as conceived in two-step frameworks of attitudes (Fazio 1990; Gawronski and Bodenhausen 2006). We thus focused on direct measures of evaluation that allow participants to exert full control over the retrieval and application of their evaluations. Theoretically and methodologically, it can be treacherous to rely on reaction-time-based indirect measures, as these measures are insensitive to negations (Deutsch et al. 2006). Hence, finding unmodulated EC effects on indirect measures could be a spurious indicator of uncontrollability. The methodological conundrum at the heart of this debate is that neither indirect nor direct measures of evaluation are process-pure reflections of automatic or nonautomatic processes. Each will reflect the contribution of both types of processes to varying degrees. Therefore, it is possible that the apparently controllable nature of direct evaluations observed in previous research masked an uncontrollable effect on some evaluations. To distinguish automatic from nonautomatic contributions to a measurement outcome, one needs to apply processing tree modeling—the primary undertaking of the current research. The Processing Tree Approach We developed an MPT model to distinguish controllable from uncontrollable EC effects in a direct evaluative measure, estimated by the model’s c- and u-parameters, respectively. Its main purpose was to investigate the potential existence of an uncontrollable EC effect on direct evaluations. A general challenge with developing a new measurement method is to demonstrate that it produces valid indicators of the targeted constructs. We have tested the validity of the method through a triangulation of approaches in our studies. Specifically, our experiments provided construct validity for our conceptualization of the c- and u-parameters as capturing an effortful, controllable process versus an efficient, uncontrollable process, respectively. Furthermore, across different EC procedures we demonstrated the external validity of our findings. We provided predictive validity by demonstrating that the parameter estimates were predictive of EC effects as assessed on independent evaluative measures as well as consumers’ choice and consumption behavior. Does the u-Parameter Reflect Uncontrolled or Uncontrollable Applications of US Valence? Readers may wonder to what degree our findings might be more reflective of the participants’ failure to exert control successfully rather than of the role of uncontrollable learning. This question may be approached from two perspectives, a conceptual and a methodological one. Conceptually, our theorizing about the role of controllable and uncontrollable processes as well as unsystematic processes is formalized in the MPT model. In that model, the failure of control (1 – c) and uncontrollable learning (u) are two orthogonal events. That is, if individuals fail to comply with task instructions with the probability 1 – c, there are two mutually exclusive outcomes. The first possible outcome is that participants applied US valence to the CS irrespective of task instructions (u). The second possible outcome is that participants did not learn uncontrollably (1 – u). In that case, they would (randomly) indicate that the CS is either positive (with the probability r) or negative (with the probability 1 – r). In other words, not following the task instructions or being incapable of doing so does not automatically lead to EC effects captured by the u-parameter. In the absence of controllable learning (1 – c), it is still important to investigate whether there is also uncontrollable learning (u). Only if a systematic process operates by which US valence is transferred to the CSs would the overall u-parameter be larger than zero. Moreover, US valence may be applied to the CSs irrespective of task instructions in an uncontrolled rather than an uncontrollable manner. For instance, participants may have lacked the capacity or motivation to exert control and may have applied US valence in a deliberate manner regardless of instructions. Therefore, it was important to investigate the u-parameter’s sensitivity to cognitive load (study 2), financial incentives (study 3), and control instructions (study 4). These experiments demonstrate the insensitivity of the u-parameter to all of these factors, demonstrating that it does not predominantly reflect deliberate processes. The present research thus attests to the partly uncontrollable nature of the EC effect. What Are the Effects of Potential Violations of the Invariance Assumption? The estimation of the MPT parameters rests on the assumption that the contributions of the latent processes are invariant across instruction conditions (Jacoby 1991). Previous research has shown that this assumption is often not met for the dominant process and that its violation biases the parameter estimation of the nondominant process (Klauer et al. 2015). In the present research, the invariance assumption for the dominant process implies that the contribution of the controllable process should be constant across standard and reversal conditions. If this assumption were violated, the estimation of the uncontrollable process would be distorted. As our designs preclude a direct, empirical test of this assumption, we cannot dismiss the possibility that the u-parameter was influenced by its violation. Yet several pieces of evidence speak against the artifactual nature of this parameter. Let us discuss one by one the two possible ways in which the contributions of the controllable process could differ between standard and reversal conditions. On the one hand, controllable processing may be instigated only by difficult tasks, in which case participants would have exerted more control in the reversal condition than in the standard condition. This notion is in line with demonstrations of increasing degrees of analytical reasoning with increasing perceptions of difficulty (Alter et al. 2007), with the conflict monitoring hypothesis that states that the degree of cognitive control is tuned to the degree of conflict (Botvinick et al. 2001), and with our observation that no control was exerted in the covert condition of study 4. If more control were exerted under reversal instructions, the contribution of the controllable processes would be overestimated, because the model would match the successfully reversed EC cases in the reversal condition with cases of CSs that acquired a regular EC effect in the standard condition. Thus, the model would treat uncontrollably acquired EC effects in the standard condition as if they were due to controllable learning. In this case, the role of uncontrollable learning would be even larger than documented in the present article. The most threatening case for the current research would arise when controllable processing plays a smaller role under reversal instructions than in the standard condition, because it could artificially inflate the u-parameter. Let us illustrate this case by means of a simplification of the modeling logic and under the assumption that there is no uncontrollable learning. If more control were exerted in the standard condition than in the reversal condition, the model would underestimate the amount of controllable learning in the standard condition, as it assumes that its contribution is as large as in the reversal condition. As this underestimation would leave unexplained cases in the standard condition that are in line with US valence, they would feed the u-parameter even though they are the result of controllable learning. Moreover, because an invariance assumption also applies to the uncontrollable component, the model would assign cases that are due only to unsystematic processes (“response tendencies” or “guessing”) in the reversal condition to the u-parameter. In this case, the u-parameter would only be an artifact reflecting a mixture of controllable processes in the standard condition and unsystematic processes in the reversal condition and have no bearing on uncontrollable evaluative learning. Several empirical findings alleviate such a concern. First, the c- and u-parameters demonstrate discriminant validity; that is, they react differently to the experimental manipulations as predicted on conceptual grounds. If the u-parameter’s estimate were dependent on the difference of control exerted in the standard versus reversal conditions, these clear-cut patterns would have been very unlikely. A second counterargument to this alternative explanation stems from the HLMs regressing the EC effect on the c- and u-parameter estimates in the within-participant designs. If the u-parameter comprised unsystematic processes in the reversal condition, it should not meaningfully predict the effects of the positive and negative stimulus pairings on independently assessed measures of the EC effect beyond the (opposite) predictive value of the c-parameter. Supporting the notion of a genuine learning process, the u-parameter predicts direction and size of the EC effect across standard and reversal conditions in a consistent manner. Moreover, its predictive value is further established for behavioral data. In summary, while we cannot exclude the possibility of violations of the invariance assumption influencing the estimate of the u-parameter, at the same time our findings speak against the notion that the u-parameter is merely an artifact of assumption violations. Theoretical Contributions Whether EC can have uncontrollable effects is part of a wider debate regarding the potential existence of automatically operating processes in associative learning. While the existence of implicitly operating processes in memory retrieval is generally uncontested, this is not the case for learning or encoding of associations (Mitchell, De Houwer, and Lovibond 2009; Shanks 2010). Despite decades of research, mostly focusing on the necessity for “contingency awareness” (another feature of automaticity), closure on this matter has not been reached (Sweldens et al. 2014). Studying the uncontrollability of evaluative learning has the potential to provide a fresh perspective on this long-standing debate. This presumes, however, that the MPT model is effective at separating nonautomatic from automatic processes. For the u-parameter in particular, this poses additional challenges. Technically, the u-parameter provided by the MPT model is an estimate of the percentage of trials where CS evaluations are in line with US valence—irrespective of the participant’s intentions or instructions—after accounting for trials that can be explained by controllable processes (which should be contingent on instructions by definition) and unsystematic processes. In other words, u is an estimate of uncontrolled CS evaluations. To make the inference leap from uncontrolled to uncontrollable, what is left to demonstrate is that this parameter is characterized by a relative independence of processing resources and motivation (Bargh 1994; Fazio 1990). Study 2 was designed to assess the effect of variations in cognitive processing capacity on the MPT parameter estimates. As expected, a reduction in cognitive resources had a detrimental effect on the likelihood to successfully exert control (c), yet the likelihood that US valence is transferred regardless of instructions remained constant and significant (u). Similarly, study 3 was designed to assess the effect of variations in motivation to exert control, by making participants’ payout contingent on their ability to follow instructions. As expected, a surge in motivation increased the likelihood that control can be successfully exerted (c). Again, however, there remained a constant likelihood that US valence would be transferred regardless of instructions (u). These data corroborate the hypothesis that the c- and u-parameters reflect the contribution of controllable and uncontrollable processes, respectively. The finding of a significant u-parameter in all our studies therefore testifies to the partially uncontrollable and automatic nature of EC effects, providing an important contribution to this debate. Substantive Implications In the US alone, companies spend $190 billion annually on advertising, a figure that continues to rise. Participants in our studies went through a series of trials in which initially neutral stimuli were paired with visual affective stimuli, resembling the experience of consumers processing a sequence of advertisements where brands or products are often linked with positive stimuli such as images and music. Our findings imply that a significant percentage of ads will change their evaluations in an uncontrollable fashion, thereby threatening their autonomy. From a consumer protection point of view, it would be vital to assess the magnitude of this effect and understand its prevalence. One powerful feature of the MPT methodology is that it provides quantitative probability estimates of nonautomatic versus automatic processes in a given setting. As revealed by the c-parameter, the ability to exert control over EC effects is far from perfect. When consumers are strongly motivated (e.g., by financial incentives in study 3), their success rate can be increased. At the same time, even a relatively simple secondary task decreased the success rate considerably (study 2). From a consumer protection point of view, this is worrying. As consumers rarely process advertisements with undivided attention, this implies they are generally unlikely to exert control. It is important to realize, however, that a failure to exert control does not automatically imply the presence of an uncontrollable EC effect. When control fails, it remains possible that no valence is transferred and the consumer’s attitude remains unchanged. However, the MPT model indicated a relatively constant likelihood that an uncontrollable EC effect would be established when control failed. It can be concluded that mere co-occurrences of brands with visual affective stimuli will often have uncontrollable effects on consumers’ evaluations and behavior despite their best efforts to correct for this influence. This could have important implications for public policy and regulation of advertising. The concern is especially acute in domains where consumers make decisions that directly affect their health and well-being. One prominent example is advertising for foods or pharmaceuticals. As outlined by Biegler and Vargas (2013), rather than being outlawed as in many other jurisdictions, in the US pharmaceutical advertising is regulated by the Food and Drug Administration, which prohibits the use of “false” or “misleading” claims in pharmaceutical messages. However, as also noted by these authors, regulations currently only verify the propositional content of these messages (textual, verifiable claims), whereas the persuasive power of these messages might rather stem from their unverifiable content (i.e., the images and musical elements), precisely the type of content our research indicates has uncontrollable effects on consumers’ attitudes. Limitations and Future Directions An important limitation of the current findings is that the exact psychological mechanisms underlying the parameters remain unclear. We obtained evidence that the process underlying the u-parameter is characterized by more features of automaticity than the drivers of the c-parameter. However, the u-parameter is still compatible with several theoretical accounts—for example, with automatically operating association formation processes (Baeyens et al. 1992) but also with implicit misattribution or “direct transfer” of evaluative responses (Jones, Fazio, and Olson 2009; Sweldens et al. 2010). One interesting direction would be to manipulate properties of the conditioning procedure, which make certain processes (e.g., implicit misattribution) more likely. Related to the previous limitation, the present research did not investigate the degree to which uncontrollable attitude acquisition shows other features of automaticity besides efficiency and unintentionality, such as unawareness or memory independence (with the latter often being used as a proxy for unawareness). We support a disjunctive view of automaticity, meaning that we do not expect all features to be either present or absent in a given EC procedure (Bargh 1994; Fiedler and Hütter 2014). Specifically, on the one hand, there may be instances where individuals are fully aware of the CS-US pairings yet are unable to control the resulting attitude. In fact, given the explicit instructions, we assume that awareness levels were very high in the present research. On the other hand, uncontrollable attitude acquisition may be facilitated by unawareness of the pairings. Such questions need to be addressed empirically. Future research should also extend our knowledge on the boundary conditions for controllable learning. For example, in the present setting participants are not provided with any information on how CSs and USs relate to each other. Moreover, participants receive no guidance on specific strategies to exert control. It would be interesting to provide consumers with different strategies to exert control and investigate their effectiveness with a sensitive MPT procedure, extending the work by Gawronski and colleagues (2015). Conclusion The method developed in this article has made it possible to test for the existence of an uncontrollable EC effect on explicit evaluations, to quantify the magnitude of its impact, and to demonstrate its downstream consequences on product choice and consumption. Thereby, this research provides an important contribution to our understanding of the acquisition of consumer attitudes. Our experiments provide evidence for independently operating controllable and uncontrollable processes. This might have important practical implications for safeguarding people’s autonomy from the influence of widespread advertising procedures, especially in sensitive domains such as food or pharmaceutical advertising. Data Collection Information The data collection for the present experiments took place between March 2012 and February 2017 at the Universities of Heidelberg, Rotterdam, and Tübingen (study 1a: March 2012, Heidelberg; study 1b: September 2013, Heidelberg; study 2: October 2012, Heidelberg; study 3: January 2014, Tübingen; study 4: February 2017, Rotterdam; study 5: May/June 2014, Tübingen). Data collection was conducted by research assistants and bachelor students (study 5) under the supervision of the first author. Data were analyzed jointly by both authors. The MPT and power analyses were conducted by Mandy Hütter. APPENDIX A: DETAILS ON STUDY MATERIALS Instructions in Studies 1a, 1b, 2 Standard Condition In the upcoming impression formation phase, you are going to learn something about persons that are yet unknown to you. In the real world, one often gets correct information from third parties about other persons. Thus, the information that will be presented during the impression formation phase in terms of pleasant and unpleasant pictures will also be informative with regard to the person depicted in the respective photograph. Reversal Condition In the upcoming impression formation phase, you are going to learn something about persons that are yet unknown to you. In the real world, one often gets incorrect information from third parties about other persons. Thus, the information that will be presented during the impression formation phase in terms of pleasant and unpleasant pictures will be informative with regard to the person depicted in the respective photograph, but it will be downright opposite to what is actually correct. That means: pleasant pictures indicate that the respective person should be shown with unpleasant pictures, while unpleasant pictures indicate that the respective person should be shown with pleasant pictures. Instructions in Study 3 Research has shown that we quickly form impressions about unknown persons and that these evaluations are heavily influenced by the context in which we see a person. Specifically, research has shown that we come to like people who appear in positive contexts, and that we come to dislike people we see in negative contexts. Whereas the context may indeed be informative about the character of a person, there are many cases where directly applying the context evaluation to person impression formation would lead to an incorrect person evaluation. For example, an evil mafia boss can be encountered on a beautiful beach. Alternatively, very nice people might be encountered in criminal neighborhoods. In this research, we investigate whether people can apply contextual information when it is accurate and whether they can reverse the influence of contextual information when it is not. Therefore, in this study we will tell you when you can trust the contextual information you see people presented in, so that it can be directly applied to make appropriate person evaluations. We will also tell you when the contextual information is actually misleading, so that you should reverse the context evaluation to make the appropriate person evaluation. In the next phase, you will therefore be presented with many different persons appearing in positive or negative contexts. You will also be informed whether the context evaluation should be directly applied to the person evaluation, or whether it should be reversed when forming an impression of the person. After the impression formation phase, we want you to report for each person whether you like her or not. We are going to count how many of your evaluations were made properly—that is, applying the context evaluation directly when it is accurate, and reversing it when it is not. Participants in the financial incentives condition then received the following additional instructions: For each appropriate classification, you will receive €0.15 as a reward. This implies that if you get them all correct, you can gain up to €3.60 on top of your base compensation (€3.00) for participation in this study. Instructions in Study 4 Standard Condition New water brands are introduced to the market on a regular basis. Previous research on advertising shows that we quickly form impressions of unknown brands and that these impressions can be strongly influenced by the context in which we encounter a brand for the first time. Specifically, this means that we may learn to like brands that appear with positive images and learn to dislike brands that appear with negative images. In advertising, such images may very well provide information about a brand. For instance, marketers might invest more money in advertising for products that have proved successful in the past or are expected to sell well. Conversely, low-quality marketing might be reflective of a low-quality product. In this research, we investigate whether people are able to APPLY the positive or negative quality of the images to the brand. In the next phase, you will therefore be presented with different water brand logos appearing repeatedly with positive or negative images. Hence, you should start LIKING the brand when paired with a POSITIVE image. Conversely, you should start DISLIKING the brand when paired with a NEGATIVE image. Afterwards, we will ask you how you feel about the different water brands. SUMMARY: If you see a POSITIVE image, you should LIKE the brand. If you see a NEGATIVE image, you should DISLIKE the brand. Reversal Condition New water brands are introduced to the market on a regular basis. Previous research on advertising shows that we quickly form impressions of unknown brands and that these impressions can be strongly influenced by the context in which we encounter a brand for the first time. Specifically, this means that we may learn to like brands that appear with positive images and learn to dislike brands that appear with negative images. While such images may well provide information about a brand in advertising, many cases are conceivable in which the direct application of the positive or negative quality of an image would lead to erroneous impressions. For example, products may be of low quality despite positive advertising. Conversely, high-quality products can be put in a negative light by competitive brands. In this research, we investigate whether people are able to REVERSE the influence of the positive or negative quality of the images to the brand. In the next phase, you will therefore be presented with different water brand logos repeatedly appearing together with positive or negative images. Hence, you should start DISLIKING the brand when paired with a POSITIVE image. Conversely, you should start LIKING the brand when paired with a NEGATIVE image. Afterwards, we will ask you how you feel about the different water brands. SUMMARY: If you see a POSITIVE image, you should DISLIKE the brand. If you see a NEGATIVE image, you should LIKE the brand. Covert Condition New water brands are introduced to the market on a regular basis. This study investigates the spontaneous reactions of consumers to different water brand logos. Since you are unlikely to have seen these brands in stores before, we will first familiarize you with them. You will see a slideshow with images not only of these water brand logos but also of landscapes and people engaged in various activities. We hope this makes the slideshow more interesting to watch. Afterwards, we will ask you how you feel about the different water brands. Please pay careful attention to this slideshow. If you miss the presentations of some water brands, you will be less familiar with them than with others and this could affect the results. SUMMARY: Please pay careful attention to this slideshow. Instructions in Study 5 Previous research on advertising shows that we quickly form impressions of unknown brands and that these impressions can be strongly influenced by the context in which we encounter a brand for the first time. Specifically, this means that we may learn to like brands that appear in positive contexts and learn to dislike brands that we encounter in negative contexts. While the context may well provide information about a brand, many cases are conceivable in which the direct application of the context would lead to erroneous impressions. For example, products may be of low quality despite positive advertising. Conversely, high-quality products can be put in a negative light by competitive brands. In this research, we investigate whether people are able to apply contextual information when it is accurate and whether they can reverse the influence of contextual information when it is not. Therefore, in this study we will tell you when you can trust the contextual information you see brands presented in, so that it can be directly applied to make appropriate brand evaluations. We will also tell you when the contextual information is actually misleading, so that you should reverse the context evaluation to make the appropriate brand evaluation. In the next phase, you will therefore be presented with many different brand logos appearing in positive or negative contexts. You will also be informed whether the context evaluation should be directly applied to evaluate the brand, or whether it should be reversed when forming an impression of the brand. After the computer-based part, participants moved to an adjacent room in the lab where the brands presented in the previous part were lined up to be tasted. Participants received the following instructions: Now that you completed all tasks on the computer, we would like to give you an opportunity to refresh yourself. These cups are filled with water of the brands you were presented with in the computer experiment. It is up to you which water brands and how many of them you would like to taste. Please decide spontaneously. We thank Olivier Corneille, the editors, and three anonymous reviewers for providing helpful feedback on previous versions of this manuscript. This research was supported by an Emmy-Noether grant (HU 1978/4-1) and a Koselleck grant (FI 294/23-1) from the German Research Foundation. We also gratefully acknowledge financial support from the INSEAD R&D and the INSEAD Alumni Fund as well as data collection assistance offered by the Erasmus Behavioral Lab at Erasmus University. APPENDIX B: FREQUENCY DATA UNDERLYING MPT ANALYSIS Observed Frequencies in the Pleasant (+) and Unpleasant (–) Response Categories as a Function of US Valence and Instruction Condition and the Respective Manipulations Standard instructions Reversal instructions + – pc + – pc Σ observations Study 1a CS+ 143 73 .66 85 143 .37 CS– 61 155 .72 124 104 .46 888 Study 1b CS+ 86 34 .72 46 74 .38 CS– 27 93 .78 80 40 .67 480 Study 2 Low load CS+ 118 62 .66 70 98 .42 CS– 51 129 .72 101 67 .40 High load CS+ 106 74 .59 85 95 .47 CS– 81 99 .55 86 94 .52 1416 Study 3 Control CS+ 154 68 .69 81 141 .36 CS– 67 155 .70 139 83 .37 Incentives CS+ 162 30 .84 41 151 .21 CS– 34 158 .82 147 45 .23 1656 Study 4 CS+ 457 203 .69 351 357 .50 CS– 274 386 .58 327 381 .54 Covert instructions CS+ 404 280 .59 CS- 318 366 .54 4104 Study 5 CS+ 406 80 .84 163 323 .34 CS- 78 408 .84 297 189 .39 1944 Standard instructions Reversal instructions + – pc + – pc Σ observations Study 1a CS+ 143 73 .66 85 143 .37 CS– 61 155 .72 124 104 .46 888 Study 1b CS+ 86 34 .72 46 74 .38 CS– 27 93 .78 80 40 .67 480 Study 2 Low load CS+ 118 62 .66 70 98 .42 CS– 51 129 .72 101 67 .40 High load CS+ 106 74 .59 85 95 .47 CS– 81 99 .55 86 94 .52 1416 Study 3 Control CS+ 154 68 .69 81 141 .36 CS– 67 155 .70 139 83 .37 Incentives CS+ 162 30 .84 41 151 .21 CS– 34 158 .82 147 45 .23 1656 Study 4 CS+ 457 203 .69 351 357 .50 CS– 274 386 .58 327 381 .54 Covert instructions CS+ 404 280 .59 CS- 318 366 .54 4104 Study 5 CS+ 406 80 .84 163 323 .34 CS- 78 408 .84 297 189 .39 1944 NOTE.— CS+: CSs paired with positive USs; CS-: CSs paired with negative USs. The proportion of US valence-congruent responses (pc) varies with instruction condition. The difference in proportions between the standard and reversal conditions can be traced back to the controllable processes leading to a reversal of CS evaluations in the reversal condition. Standard instructions Reversal instructions + – pc + – pc Σ observations Study 1a CS+ 143 73 .66 85 143 .37 CS– 61 155 .72 124 104 .46 888 Study 1b CS+ 86 34 .72 46 74 .38 CS– 27 93 .78 80 40 .67 480 Study 2 Low load CS+ 118 62 .66 70 98 .42 CS– 51 129 .72 101 67 .40 High load CS+ 106 74 .59 85 95 .47 CS– 81 99 .55 86 94 .52 1416 Study 3 Control CS+ 154 68 .69 81 141 .36 CS– 67 155 .70 139 83 .37 Incentives CS+ 162 30 .84 41 151 .21 CS– 34 158 .82 147 45 .23 1656 Study 4 CS+ 457 203 .69 351 357 .50 CS– 274 386 .58 327 381 .54 Covert instructions CS+ 404 280 .59 CS- 318 366 .54 4104 Study 5 CS+ 406 80 .84 163 323 .34 CS- 78 408 .84 297 189 .39 1944 Standard instructions Reversal instructions + – pc + – pc Σ observations Study 1a CS+ 143 73 .66 85 143 .37 CS– 61 155 .72 124 104 .46 888 Study 1b CS+ 86 34 .72 46 74 .38 CS– 27 93 .78 80 40 .67 480 Study 2 Low load CS+ 118 62 .66 70 98 .42 CS– 51 129 .72 101 67 .40 High load CS+ 106 74 .59 85 95 .47 CS– 81 99 .55 86 94 .52 1416 Study 3 Control CS+ 154 68 .69 81 141 .36 CS– 67 155 .70 139 83 .37 Incentives CS+ 162 30 .84 41 151 .21 CS– 34 158 .82 147 45 .23 1656 Study 4 CS+ 457 203 .69 351 357 .50 CS– 274 386 .58 327 381 .54 Covert instructions CS+ 404 280 .59 CS- 318 366 .54 4104 Study 5 CS+ 406 80 .84 163 323 .34 CS- 78 408 .84 297 189 .39 1944 NOTE.— CS+: CSs paired with positive USs; CS-: CSs paired with negative USs. The proportion of US valence-congruent responses (pc) varies with instruction condition. The difference in proportions between the standard and reversal conditions can be traced back to the controllable processes leading to a reversal of CS evaluations in the reversal condition. REFERENCES Allen Chris T. , Madden Thomas J. ( 1985 ), “ A Closer Look at Classical Conditioning,” Journal of Consumer Research , 12 3 , 301 – 15 . Google Scholar CrossRef Search ADS Alter Adam L. , Oppenheimer Daniel M. , Epley Nicholas , Eyre Rebecca N. ( 2007 ), “Overcoming Intuition: Metacognitive Difficulty Activates Analytic Reasoning,” Journal of Experimental Psychology: General , 136 4 , 569 – 76 . Google Scholar CrossRef Search ADS PubMed Baeyens Frank , Eelen Paul , Crombez Geert , Vandenbergh Omer ( 1992 ), “ Human Evaluative Conditioning: Acquisition Trials, Presentation Schedule, Evaluative Style and Contingency Awareness,” Behaviour Research and Therapy , 30 2 , 133 – 42 . Google Scholar CrossRef Search ADS PubMed Bargh John A. ( 1994 ), “The Four Horsemen of Automaticity: Awareness, Intention, Efficiency, and Control in Social Cognition,” in Handbook of Social Cognition, Vol. 1: Basic Processes , 2nd ed. , ed. Wyer Robert S. Jr. , Srull Thomas K. , Hillsdale, NJ : Erlbaum , 1 – 40 . Batchelder William H. , Riefer David M. ( 1999 ), “ Theoretical and Empirical Review of Multinomial Process Tree Modeling,” Psychonomic Bulletin & Review , 6 1 , 57 – 86 . Google Scholar CrossRef Search ADS PubMed Biegler Paul , Vargas Patrick ( 2013 ), “ Ban the Sunset? Nonpropositional Content and Regulation of Pharmaceutical Advertising,” American Journal of Bioethics , 13 5 , 3 – 13 . Google Scholar CrossRef Search ADS PubMed Bierley Calvin , McSweeney Frances K. , Vannieuwkerk Renee ( 1985 ), “ Classical Conditioning of Preferences for Stimuli,” Journal of Consumer Research , 12 3 , 316 – 23 . Google Scholar CrossRef Search ADS Botvinick Matthew M. , Braver Todd S. , Barch Deanna M. , Carter Cameron S. , Cohen Jonathan D. ( 2001 ), “Conflict Monitoring and Cognitive Control,” Psychological Review , 108 3 , 624 – 52 . Google Scholar CrossRef Search ADS PubMed Cohen Jacob ( 1988 ), Statistical Power Analysis for the Behavioral Sciences , Hillsdale, NJ : Erlbaum . Corneille Olivier , Yzerbyt Vincent , Pleyers Gordy , Mussweiler Thomas ( 2009 ), “Beyond Awareness and Resources: Evaluative Conditioning May Be Sensitive to Processing Goals,” Journal of Experimental Social Psychology , 45 1 , 279 – 82 . Google Scholar CrossRef Search ADS Curran Tim , Hintzman Douglas L. ( 1995 ), “Violations of the Independence Assumption in Process Dissociation,” Journal of Experimental Psychology: Learning, Memory, and Cognition , 21 3 , 531 – 47 . Google Scholar CrossRef Search ADS PubMed Dalton Amy N. , Huang Li ( 2014 ), “ Motivated Forgetting in Response to Social Identity Threat,” Journal of Consumer Research , 40 6 , 1017 – 38 . Google Scholar CrossRef Search ADS De Houwer Jan ( 2006 ), “Using the Implicit Association Test Does Not Rule Out an Impact of Conscious Propositional Knowledge on Evaluative Conditioning,” Learning and Motivation , 37 2 , 176 – 87 . Google Scholar CrossRef Search ADS De Houwer Jan , Teige-Mocigemba Sarah , Spruyt Adriaan , Moors Agnes ( 2009 ), “ Implicit Measures: A Normative Analysis and Review,” Psychological Bulletin , 135 3 , 347 – 68 . Google Scholar CrossRef Search ADS PubMed De Houwer Jan , Thomas Sarah , Baeyens Frank ( 2001 ), “Associative Learning of Likes and Dislikes: A Review of 25 Years of Research on Human Evaluative Conditioning,” Psychological Bulletin , 127 6 , 853 – 69 . Google Scholar CrossRef Search ADS PubMed Dedonder Jonathan , Corneille Olivier , Yzerbyt Vincent , Kuppens Toon ( 2010 ), “Evaluative Conditioning of High-Novelty Stimuli Does Not Seem to Be Based on an Automatic Form of Associative Learning,” Journal of Experimental Social Psychology , 46 6 , 1118 – 21 . Google Scholar CrossRef Search ADS Deutsch Roland , Gawronski Bertram , Strack Fritz ( 2006 ), “ At the Boundaries of Automaticity: Negation as Reflective Operation,” Journal of Personality and Social Psychology , 91 3 , 385 – 405 . Google Scholar CrossRef Search ADS PubMed Deutsch Roland , Kordts-Freudinger Robert , Gawronski Bertram , Strack Fritz ( 2009 ), “Fast and Fragile: A New Look at the Automaticity of Negation Processing,” Experimental Psychology , 56 6 , 434 – 46 . Google Scholar CrossRef Search ADS PubMed Fazio Russell H. ( 1990 ), “Multiple Processes by Which Attitudes Guide Behavior: The MODE Model as an Integrative Framework,” in Experimental Social Psychology , Vol. 23 , ed. Zanna Mark P. , San Diego, CA : Academic Press , 75 – 109 . Fiedler Klaus , Hütter Mandy ( 2014 ), “The Limits of Automaticity,” in Dual-Process Theories of the Social Mind , ed. Sherman Jeffrey , Gawronski Bertram , Trope Yaacov , New York : Guilford , 497 – 513 . Fiedler Klaus , Unkelbach Christian ( 2011 ), “Evaluative Conditioning Depends on Higher Order Cognitive Processes,” Cognition & Emotion , 25 4 , 639 – 56 . Google Scholar CrossRef Search ADS PubMed Fitzsimons Gavan J. , Williams Patti ( 2000 ), “ Asking Questions Can Change Choice Behavior: Does It Do So Automatically or Effortfully?” Journal of Experimental Psychology: Applied , 6 3 , 195 – 206 . Google Scholar CrossRef Search ADS PubMed Gawronski Bertram , Balas Robert , Creighton Laura A. ( 2014 ), “ Can the Formation of Conditioned Attitudes Be Intentionally Controlled?” Personality and Social Psychology Bulletin , 40 4 , 419 – 32 . Google Scholar CrossRef Search ADS PubMed Gawronski Bertram , Bodenhausen Galen V. ( 2006 ), “ Associative and Propositional Processes in Evaluation: An Integrative Review of Implicit and Explicit Attitude Change,” Psychological Bulletin , 132 5 , 692 – 731 . Google Scholar CrossRef Search ADS PubMed Gawronski Bertram , Mitchell Derek G. , Balas Robert ( 2015 ), “ Is Evaluative Conditioning Really Uncontrollable? A Comparative Test of Three Emotion-Focused Strategies to Prevent the Acquisition of Conditioned Preferences,” Emotion , 15 5 , 556 – 68 . Google Scholar CrossRef Search ADS PubMed Gibson Bryan ( 2008 ), “ Can Evaluative Conditioning Change Attitudes Toward Mature Brands? New Evidence from the Implicit Association Test,” Journal of Consumer Research , 35 1 , 178 – 88 . Google Scholar CrossRef Search ADS Gorn Gerald J. ( 1982 ), “ The Effects of Music in Advertising on Choice Behavior: A Classical Conditioning Approach,” Journal of Marketing , 46 1 , 94 – 101 . Google Scholar CrossRef Search ADS Hofmann Wilhelm , De Houwer Jan , Perugini Marco , Baeyens Frank , Crombez Geert ( 2010 ), “ Evaluative Conditioning in Humans: A Meta-Analysis,” Psychological Bulletin , 136 3 , 390 – 421 . Google Scholar CrossRef Search ADS PubMed Hu Xiangen , Batchelder William H. ( 1994 ), “The Statistical Analysis of Engineering Processing Tree Models with the EM Algorithm,” Psychometrika , 59 1 , 21 – 47 . Google Scholar CrossRef Search ADS Hütter Mandy , Klauer Karl C. ( 2016 ), “Applying Processing Trees in Social Psychology,” European Review of Social Psychology , 27 1 , 116 – 59 . Google Scholar CrossRef Search ADS Hütter Mandy , Sweldens Steven ( 2013 ), “Implicit Misattribution of Evaluative Responses: Contingency-Unaware Evaluative Conditioning Requires Simultaneous Stimulus Presentations,” Journal of Experimental Psychology: General , 142 3 , 638 – 43 . Google Scholar CrossRef Search ADS PubMed Hütter Mandy , Sweldens Steven , Stahl Christoph , Unkelbach Christian , Klauer Karl C. ( 2012 ), “Dissociating Contingency Awareness and Conditioned Attitudes: Evidence of Contingency-Unaware Evaluative Conditioning,” Journal of Experimental Psychology: General , 141 3 , 539 – 57 . Google Scholar CrossRef Search ADS PubMed Ito Tiffany A. , Larsen Jeff T. , Smith N. Kyle , Cacioppo John T. ( 1998 ), “ Negative Information Weighs More Heavily on the Brain: The Negativity Bias in Evaluative Categorizations,” Journal of Personality and Social Psychology , 75 4 , 887 – 900 . Google Scholar CrossRef Search ADS PubMed Jacoby Larry L. ( 1991 ), “A Process Dissociation Framework—Separating Automatic from Intentional Uses of Memory,” Journal of Memory and Language , 30 5 , 513 – 41 . Google Scholar CrossRef Search ADS Jacoby Larry L. , Shrout Patrick E. ( 1997 ), “Toward a Psychometric Analysis of Violations of the Independence Assumption in Process Dissociation,” Journal of Experimental Psychology: Learning, Memory, and Cognition , 23 2 , 505 – 10 . Google Scholar CrossRef Search ADS Janiszewski Chris , Warlop Luk ( 1993 ), “ The Influence of Classical Conditioning Procedures on Subsequent Attention to the Conditioned Brand,” Journal of Consumer Research , 20 2 , 171 – 89 . Google Scholar CrossRef Search ADS Johar Gita V. , Maheswaran Durairaj , Peracchio Laura A. ( 2006 ), “Mapping the Frontiers: Theoretical Advances in Consumer Research on Memory, Affect, and Persuasion,” Journal of Consumer Research , 33 1 , 139 – 49 . Google Scholar CrossRef Search ADS Jones Christopher R. , Fazio Russell H. , Olson Michael A. ( 2009 ), “Implicit Misattribution as a Mechanism Underlying Evaluative Conditioning,” Journal of Personality and Social Psychology , 96 5 , 933 – 48 . Google Scholar CrossRef Search ADS PubMed Klauer Karl Christoph , Dittrich Kerstin , Scholtes Christine , Voss Andreas ( 2015 ), “The Invariance Assumption in Process-Dissociation Models: An Evaluation Across Three Domains,” Journal of Experimental Psychology: General , 144 1 , 198 – 221 . Google Scholar CrossRef Search ADS PubMed Lang Peter J. , Bradley Margaret M. , Cuthbert Bruce N. ( 2008 ), International Affective Picture System (IAPS): Affective Ratings of Pictures and Instruction Manual . Technical Report A-8, Gainesville : University of Florida . Mierop Adrien , Hütter Mandy , Corneille Olivier ( 2017 ), “Resource Availability and Explicit Memory Largely Determine Evaluative Conditioning Effects in a Paradigm Claimed to Be Conducive to Implicit Attitudes Acquisition,” Social Psychological and Personality Science , 8 7 , 758 – 67 . Google Scholar CrossRef Search ADS Mitchell Chris J. , De Houwer Jan , Lovibond Peter F. ( 2009 ), “ The Propositional Nature of Human Associative Learning,” Behavioral and Brain Sciences , 32 2 , 183 – 98 . Google Scholar CrossRef Search ADS PubMed Moors Agnes , De Houwer Jan ( 2006 ), “Automaticity: A Conceptual and Theoretical Analysis,” Psychological Bulletin , 132 2 , 297 – 326 . Google Scholar CrossRef Search ADS PubMed Moshagen Morten ( 2010 ), “MultiTree: A Computer Program for the Analysis of Multinomial Processing Tree Models,” Behavior Research Methods , 42 1 , 42 – 54 . Google Scholar CrossRef Search ADS PubMed Olson Michael A. , Fazio Russell H. ( 2001 ), “ Implicit Attitude Formation through Classical Conditioning,” Psychological Science , 12 5 , 413 – 7 . Google Scholar CrossRef Search ADS PubMed Page Monte M. ( 1969 ), “Social Psychology of a Classical Conditioning of Attitudes Experiment,” Journal of Personality and Social Psychology , 11 2 , 177 – 86 . Google Scholar CrossRef Search ADS Payne B. Keith , Bishara Anthony J. ( 2009 ), “An Integrative Review of Process Dissociation and Related Models in Social Cognition,” European Review of Social Psychology , 20 1 , 272 – 314 . Google Scholar CrossRef Search ADS Pham Michel T. , Johar Gita V. ( 1997 ), “ Contingent Processes of Source Identification,” Journal of Consumer Research , 24 3 , 249 – 65 . Google Scholar CrossRef Search ADS Pleyers Gordy , Corneille Olivier , Yzerbyt Vincent , Luminet Olivier ( 2009 ), “Evaluative Conditioning May Incur Attentional Costs,” Journal of Experimental Psychology: Animal Behavior Processes , 35 2 , 279 – 85 . Google Scholar CrossRef Search ADS PubMed Rogers Robert D. , Monsell Stephen ( 1995 ), “Costs of a Predictable Switch between Simple Cognitive Tasks,” Journal of Experimental Psychology: General , 124 , 207 – 31 . Google Scholar CrossRef Search ADS Shanks David R. ( 2010 ), “ Learning: From Association to Action,” Annual Review of Psychology , 61 , 273 – 301 . Google Scholar CrossRef Search ADS PubMed Shapiro Stewart ( 1999 ), “When an Ad’s Influence Is Beyond Our Conscious Control: Perceptual and Conceptual Fluency Effects Caused by Incidental Ad Exposure,” Journal of Consumer Research , 26 1 , 16 – 36 . Google Scholar CrossRef Search ADS Shimp Terence A. , Stuart Elnora W. , Engle Randall W. ( 1991 ), “ A Program of Classical-Conditioning Experiments Testing Variations in the Conditioned-Stimulus and Context,” Journal of Consumer Research , 18 1 , 1 – 12 . Google Scholar CrossRef Search ADS Stahl Christoph , Klauer Karl C. ( 2007 ), “HMMTree: A Computer Program for Latent-Class Hierarchical Multinomial Processing Tree Models,” Behavior Research Methods , 39 2 , 267 – 73 . Google Scholar CrossRef Search ADS PubMed Stuart Elnora W. , Shimp Terence A. , Engle Randall W. ( 1987 ), “ Classical Conditioning of Consumer Attitudes: Four Experiments in an Advertising Context,” Journal of Consumer Research , 14 3 , 334 – 49 . Google Scholar CrossRef Search ADS Sweldens Steven , Corneille Olivier , Yzerbyt Vincent ( 2014 ), “ The Role of Awareness in Attitude Formation through Evaluative Conditioning,” Personality and Social Psychology Review , 18 2 , 187 – 209 . Google Scholar CrossRef Search ADS PubMed Sweldens Steven , Tuk Mirjam , Hütter Mandy ( 2017 ), “How to Study Consciousness in Consumer Research: A Commentary on Williams and Poehlman,” Journal of Consumer Research , 44 2 , 266 – 75 . Sweldens Steven , van Osselaer Stijn M. J. , Janiszewski Chris ( 2010 ), “ Evaluative Conditioning Procedures and the Resilience of Conditioned Brand Attitudes,” Journal of Consumer Research , 37 3 , 473 – 89 . Google Scholar CrossRef Search ADS Tukey John W. ( 1977 ), Exploratory Data Analysis, Reading, PA : Addison-Wesley . Wegner Daniel M. ( 1994 ), “Ironic Processes of Mental Control,” Psychological Review , 101 1 , 34 – 52 . Google Scholar CrossRef Search ADS PubMed Whitmer Anson J. , Banich Marie T. ( 2007 ), “Inhibition versus Switching Deficits in Different Forms of Rumination,” Psychological Science , 18 6 , 546 – 53 . Google Scholar CrossRef Search ADS PubMed Williams Lawrence E. , Poehlman Andrew ( 2017 ), “Conceptualizing Consciousness in Consumer Research,” Journal of Consumer Research , 44 2 , 231 – 51 . © The Author(s) 2018. Published by Oxford University Press on behalf of Journal of Consumer Research, Inc. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)

Journal

Journal of Consumer ResearchOxford University Press

Published: Feb 2, 2018

There are no references for this article.

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off