functional interactions between the environment and behaviour, the current data may also be of interest to researchers operating at the mental (mechanistic) level of analysis. As outlined in Chapter 2, three mental models currently dominate discourse within this research area and accommodate our findings with differing levels of success. We will now consider each of these models in turn.
Associative Mental Models
The above results appear to seriously conflict with associative accounts of evaluative responding (Baeyens et al., 1992; Martin & Levey, 1994). According to this perspective, the direct pairing of stimuli leads to the formation of unqualified links between mental representations in memory. Thus presenting one stimulus with positive images and another with negative images should result in those stimuli being evaluated in-line with the valence implied by the images they were paired with – regardless of the relation established between those stimuli by a contextual cue. In Experiments 1-3, for example, relating Pokémon-Same-
Positive and Pokémon-Opposite-Positive should have resulted in identical (positive)
outcomes on the various measures of evaluation. Similarly, presenting Pardal with positive or Ciney with negative words and reinforcing the selection of the ‘Opposite’ cue should have
153
resulted in the former brand being rated positively and the latter negatively (Experiment 4). None of these predictions were confirmed in any of the above studies, with the direction - and in the case of Experiments 5 and 6 - magnitude of evaluative responding dictated by the cue governing the stimulus relation.
Associative accounts also face the additional challenge of explaining how stimuli that were never paired with valenced words or images came to elicit responses on direct and indirect procedures. In Experiments 2-3, for instance, Pokémon 3 and 6 should not have elicited evaluative responses given that neither stimulus was directly paired with valenced stimuli nor were they paired with Pokémon 1 or 4 during the task. Likewise, the brands Ettalas and Gageleer in Experiment 4 and the prize Ciney in Experiments 5 and 6 should not have been liked or disliked for the very same reason. Even allowing for some combination of higher-order backward sensory preconditioning (see above) these models fail to explain why the psychological functions established for one stimulus were not simply transferred but rather transformed through the derived relations. For instance, when Pokémon 1 was related with negative images using the ‘Opposite’ cue participants liked Pokémon 2 and 3 (Experiment 3); when Pardal was Opposite-Positive both Zatte and Ettalas were disliked (Experiment 4) while Ciney was liked more than Ettalas or Zatte in Experiments 5-6. Finally, the fact that indirect effects only emerged when participants correctly derived the relation between stimuli seems to provide the strictest challenge for this class of models. Given that implicit or automatic evaluations are often assumed to be governed by associations in memory formed on the basis of experienced pairings (Gawronski, Deutsch & Banse, 2011; Nosek et al., 2011), IAT, IRAP and affective priming effects should not have been observed at any point in the current thesis. Contrary to this claim, participants displayed evidence of derived coordination, opposition and comparative relating across a variety of indirect
154
procedures and stimulus sets. As such, traditional associative explanations appear to be problematic on multiple fronts.
Propositional Models
The current data not only undermine associative models but provide firm support for propositional accounts that involve qualified links between mental representations in memory. From this perspective, changes in evaluative responding are moderated by direct and derived stimulus relating which is in turn mediated by the formation of propositions concerning those relations. Whereas associations simply convey the strength with which representations are linked in memory, propositions specify their strength, structure and content (e.g., “X is opposite to Y” or “X is larger than Y”). Likewise, while associations gradually develop with many experienced pairings (Smith & DeCoster, 2000; Strack & Deutsch, 2004), propositions can be formed on the basis of direct training or inferred via deductive reasoning and language (De Houwer, 2009a). When combined these two properties of propositions may explain the current set of findings.
In our studies participants were exposed to contextual cue training in which the selection of an arbitrary symbol was reinforced in the presence of pictures bearing a relation of similarity or opposition. Across numerous different exemplars they abstracted out the relational functions of ‘Same’ for one cue and ‘Opposite’ for the other. From a mechanistic point of view, this history of relational learning could be interpreted as one that gave rise to ‘Same’ and ‘Opposite’ propositions about the two arbitrary symbols (e.g., “This symbol
means that the pictures are the same and that symbol means they are opposite”). Thereafter,
a series of stimulus relations were formed by differentially reinforcing the selection of one of these symbols in the presence of a stimulus combination. Mechanistically speaking, the differential selection of the ‘Same’ and ‘Opposite’ cues across stimulus pairs may have resulted in the formation of additional propositions about the trained relations (e.g.,
155
“Pokémon 1 is the same as Pokémon 2”, “Pokémon 2 is the same as Pokémon 3” and “Pokémon 1 is the opposite of negative images”). Critically, participants who passed the derivation test may have also made a “propositional leap” insofar as they generated a set of novel, untrained propositions about the derived relations (e.g., “Pokémon 3 is good and
Pokémon 6 is bad”). If this latter assumption is correct, then it was these inferred propositions
(rather than their directly trained counterparts) that mediated responding on the direct and indirect procedures. This may help to explain why only participants who successfully derived (i.e., made a “propositional leap”) showed evidence for evaluative responding while their counterparts who failed that test (but passed all stages of contextual and relational training) did not. In other words, a series of propositions based on direct stimulus relating may not have been enough – rather these propositions may have to give rise to additional propositions in order to produce the expected effects.
Proponents of propositional models often argue that the formation of these mental constructs requires not only an awareness of the to-be-related stimuli but the cognitive resources, time and goals to relate them (De Houwer, 2009a). Although we did not manipulate these various mental conditions in the current thesis (given our functional perspective) it could be argued that they were nevertheless present during contextual cue and relational training. For instance, progression through training and testing trials was free from time constraint, required 100% accuracy and did not involve a distractor or non-relevant task during learning (i.e., training was non-time pressured, goal directed, and explicitly required an awareness of the stimuli). While admittedly speculative, mechanistic researchers could replicate the current work while controlling for the above factors. Our findings also introduce the possibility that once formed propositions may be activated automatically from memory and guide the evaluation of stimuli on direct and indirect measures alike (see Hughes, Barnes-Holmes & De Houwer, 2011). For example, once the proposition “Ciney is more than
156
Pardal” has been generated based on the derived relation between those stimuli, this
proposition may have been stored in memory. In contrast to dual-process accounts (see
below) this memory representation is assumed to be propositional in nature. Although
mechanistic researchers have often appealed to mental associations in explaining indirect task performances, there is no a priori reason why propositional knowledge could not be activated from memory and lead to automatic evaluations of a given stimulus.
Dual-Process Models
Finally, dual-process accounts that allow for rules (Smith & DeCoster, 2000), judgments (Kahneman, 2003), and propositions (Gawronski & Bodenhausen, 2011) to feed into and create novel associations could account for our findings – provided that certain pre- conditions are met. For instance, and similar to the above propositional model, contextual cue training may be mediated by the formation of propositions concerning the meaning of the two symbols. These propositions could then be used during relational training to generate additional propositions about the stimulus relations themselves. This set of “directly trained” propositions could subsequently give rise to novel, untrained propositions about stimuli that were never paired together. If correct then these inferred propositions could be transformed into and stored as associations in memory, explaining the direct and indirect outcomes observed in Experiments 1-6.
On balance, while interactive dual-process models may accommodate our transformation of function effects they do not appear to do so in an a priori fashion. Rather it appears that this particular class of models can account for virtually every empirical finding post-hoc. In addition, they often fail to provide clear testable predictions for their falsification or even offer a means to empirically distinguish between the explanatory concepts of one dual-process model relative to another (for a discussion see Gawronski & Creighton, in press; Keren & Schul, 2009). Our findings thus pose a challenge to dual-process (and propositional)
157
model theorists to articulate how, when and why (a) the contextual cue and relational training procedures give rise to propositions (b) novel propositions will emerge based on a set of directly trained propositions, and in the case of dual-process accounts, (c) how these propositions are subsequently transformed into and stored as associations given that participants were never exposed to any training or instructions to do so. More generally, while the encoding of affirmation or negation propositions seems to fit with traditional notions of mental associations (e.g., “Pokémon-Good”) it is not immediately clear how other propositions, such as those involving comparison, distinction, or hierarchy are stored as unqualified links in memory given their inherently relational content.
Summary. In short, mechanistic researchers may explain the current results as being
moderated by derived stimulus relations encountered in the environment and mediated by the
formation of either mental associations, propositions or some combination of the two. Within this world-view theoretical models are usually evaluated on two fronts - their ability to account for existing knowledge about the phenomenon of interest (heuristic value) and ability to make a priori predictions rather than post-hoc rationalizations (predictive value). Given that evaluative responding only emerged when participants correctly derived the relation between stimuli, and that derivation can conceptually be achieved on the basis of propositions but not associations, the current results provide strong support for propositional (and to a lesser extent dual-process) models and against their purely associative counterparts. It must be said, however, that single process models involving propositions seem to provide a coherent and parsimonious explanation for the above data without the need for recourse to a second (associative) mental construct.