discussion.Rmd

# General Discussion

We begin by briefly summarizing the findings of the present study. 
Next, we turn to its limitations and identify open questions for future research.

## Main findings

Across a series of experiments, we manipulated, within-subjects, the identifiability of CS stimuli by impairing stimulus strength. 
When CSs were presented for 30 ms and were masked, identification was low but remained clearly above chance levels. 
When CSs were either not masked, or presented for 100 ms or longer, they were clearly identifiable.
These findings suggest that participants can partially identify CS stimuli even under brief and masked presentation conditions.

EC effects were found in Experiments 2-5 for CS that were clearly identifiable -- either because they were presented for a longer duration (100ms in Experiment 4; 1000ms in Experiment 5), or because they were not masked (Experiments 2 & 3). 
Despite above-chance levels of CS identification, no EC effects were found under brief and masked presentation conditions in Experiments 1-6. 
This finding was replicated with different methods of reducing stimulus strength (i.e., masking in Exp. 2 & 3; presentation duration in Expts. 1, 4-6); 
it did not depend on the requirement of trial-wise identification responses (Exp. 2); 
it was replicated across different orientation tasks (valence judgments in Expts. 2 & 3; brightness judgments in Expts. 3 & 4; CS & US identification in Exp. 6; CS identification in Expts. 1 & 5); 
it was unaffected by the response requirement in the orienting task (Exp. 4);
and it was replicated across different types of conditioned stimuli (nonwords, faces, products).
These results suggest that EC does not occur with briefly presented and pattern-masked CSs.

Taken together, when we experimentally reduced CS stimulus strength to low but still above-chance levels of CS identification, EC effects were eliminated.
In other words, we found that supraliminal presentation of the CS was necessary (but not sufficient) for EC in a logical sense, such that EC did not occur for subliminal CSs. 
This suggests that awareness is necessary (but not sufficient) for EC in a logical sense [@dawson_role_1976].
The present findings do not, however, support the claim that awareness plays a causal role in EC: 
As we have only indirectly manipulated awareness via stimulus strength, the present study is not capable of addressing the causal status, or mediating role, of awareness.
Nevertheless, the present findings are of course consistent both with strong as well as weak single-process learning models [@lovibond_role_2002]:
In the strong model, a single (propositional) learning process gives rise to awareness, which then plays a causal role in the production of the learning phenomenon [@mitchell_propositional_2009].
In the weak model, the learning process gives rise to both awareness and learning; in this model, although correlated with learning, awareness has no causal role and is merely epiphenomenal.
The present finding is, however, inconsistent with dual-process models postulating awareness-independent evaluative learning [e.g., @gawronski_associative_2006; @jones_implicit_2009].



## Additional findings

### The role of orienting task

Across studies, we manipulated the orienting task participants performed during the learning phase. 
On the one hand, EC was not robustly found -- even for clearly supraliminal CSs presented for 100ms -- when the orienting task induced a strong focus on CS identification and allowed participants to ignore the US (although EC was obtained with this orienting task in Exp. 5 with CSs presented for 1000ms). 
On the other hand, EC was obtained when participants were instructed to judge the valence of the CS-US pair, when they were instructed to judge the brightness of the CS-US pair, or when they were asked to attend to both the CS and the US. 
The orienting tasks differ with regard to several features that differ in their ability to explain this pattern.

In the EC literature, two types of associations are discussed as potentially underlying EC effects. First, referential learning accounts assume that CS-US (or S-S) associations are learned, such that retrieval of the CS activates the representation of the US, thereby allowing for its valence to affect responses. In contrast, holistic (S-R) accounts assume that during learning, the US elicits an evaluative response that is then associated with the CS; later CS retrieval then activates the associated evaluative response. Both the valence-focus orienting task as well as the surveillance task have been argued to favor S-R learning [although different processes have been postulated for each task; see @gast_i_2011; @jones_implicit_2009]; it is not known whether the brightness-focus task favors one of the two types of associations.
^[In addition, S-R learning has been claimed to operate when CSs are paired with multiple different USs of the same valence and when there is temporal overlap between CS and US presentations [@sweldens_evaluative_2010]; both conditions were realized in the present studies.] 
Because S-R learning is assumed to underlie EC effects in both the valence-focus task (which yielded large EC effects) and the surveillance task (which yielded small effects), the type of association – S-S versus S-R -- is unlikely to account for the differences in EC effect size obtained across the present experiments.

The pattern of EC effects across orientation tasks may more readily be interpreted as suggesting that processing mode modulates EC, with an integrative or holistic mode presumably induced by a valence or brightness task leading to greater EC effects.
In a similar vein, previous work has shown that EC was greater when participants processed similarities between CS and US than when they focused on differences [@corneille_beyond_2009].
The fact that, in other studies, EC has been found without special orienting instructions may be explained by the assumption that a holistic mode is the default. 
Yet, an explanation in terms of processing mode appears to be inconsistent with the finding of EC effects in a surveillance orienting task [@olson_implicit_2001] in which participants searched for a specific target stimulus, suggesting an analytic mode. 
Thus, it remains to be seen whether the moderating variable underlying the present pattern is indeed processing mode.

The above findings could also be explained by CS-US integration. 
They are consistent with recent work suggesting that a relational or integrative processing of CS and US is beneficial for EC [e.g., @jones_implicit_2009]. 
If EC depends on the formation of an association between CS and US in memory, then such an association should be more strongly formed when CS and US are both attended to during learning and are therefore more likely to be stored in an integrated trace or episode [see also @kattner_revisiting_2012].

Another interpretation is in terms of attention to the information relevant for EC, that is, the CS-US pair and its valence. 
Across studies, we implemented four different levels of this variable: 
The highest level was present in the valence-focus task in which participants were instructed to attend to the CS-US pair and to judge its valence;
EC effects were greatest under these conditions ($d=`r metaEC.supra_val$TE.random`$, 95% CI: $`r metaEC.supra_val$lower.random`, `r metaEC.supra_val$upper.random`$).
Attention to the relevant information was reduced in the brightness-focus condition in which participants were also instructed to attend to the CS-US pair, but were asked to focus on brightness instead of valence.
EC effects under these brightness-focus conditions tended to be markedly smaller than those obtained under valence-focus conditions ($d=`r metaEC.supra_bri$TE.random`$, 95% CI: $`r metaEC.supra_bri$lower.random`, `r metaEC.supra_bri$upper.random`$). 
We further reduced attention to the relevant information in the last experiment in which participants were instructed to attend to the identity of the individual stimuli (i.e., both CS and US), but were not instructed to process these stimuli as CS-US *pairs*; 
only small EC effects were obtained under these conditions ($d=`r metaEC.supra_csus$TE.random`$, 95% CI: $`r metaEC.supra_csus$lower.random`, `r metaEC.supra_csus$upper.random`$). 
Finally, EC effects were also small (even undetectable in some individual studies) when participants were instructed to attend to only the CS and encouraged to ignore the USs (which were introduced as background images; $d=`r metaEC.supra_cs$TE.random`$, 95% CI: $`r metaEC.supra_cs$lower.random`, `r metaEC.supra_cs$upper.random`$).
These findings suggest that the differences between orienting tasks may be interpreted as a function of the attention they direct toward the CS-US pair and its valence.

### Main effects of masking and presentation duration on CS evaluation

Across studies, CSs were evaluated more positively when presented for a longer time or when they were not masked. 
We surmise that this finding may be interpreted as an evaluative conditioning effect.
Specifically, longer or unmasked CS presentations may have elicited more positive feelings because of their enhanced ease of detection. 
Conversely, brief and masked CS may have elicited some frustration, resulting in relatively more negative evaluations. 
If this interpretation is correct, it would suggest that
(1) participants’ metacognitive experiences during their processing of a CS-US pairing may act as a US; 
(2) such experiences are influential enough to add to effects by the US stimuli presented together with the CS; and 
(3) this effect may occur spontaneously (as it was not moderated by the presence or absence of an identification task in Experiment 2). 
The latter hypothesis is consistent with other metacognition-driven evaluative learning effects reported in the literature. 
For instance, people devaluated stimuli that distracted them in a previous selective attention task [@fenske_affective_2006]. 
In addition, preferences increased for objects that were actively gazed at [@shimojo_gaze_2003]. 
And, perhaps conceptually closer to our current findings, it has been argued that mere exposure should be considered a special case of evaluative conditioning [@zajonc_mere_2001], in which the absence of noxious consequences of a repeatedly exposed stimulus (i.e., the CS) would act as an “internal” US that increases the CS liking. 
That both external and internal stimuli can result in evaluative conditioning effects is also consistent with instructed evaluative conditioning effects, in which participants are asked to categorize the CSs either positively or negatively [@gast_evaluative_2012].

 


## Limitations and open questions

The present studies have implemented a commonly used procedure of investigating subliminal processes -- pattern masking -- to study EC effects for subliminal CSs.
Under these specific conditions, we did not find such subliminal EC effects.
It remains possible that subliminal EC effects will be found under different presentation conditions.
In this section we briefly discuss conditions we deem relevant for investigation in future research.


### EC effects for briefly presented stimuli may have failed to obtain because implicit misattribution operated on the mask stimuli (instead of the CSs)

Because the joint presentation of CS and US was much briefer (i.e., 30 ms) than the joint presentation of the US and the forward and backward masks (i.e., 500 ms and 1470 ms), an association may have formed only for the mask stimuli but not for the CS. 
For instance, in terms of affect misattribution, an affective response elicited by the US may have been misattributed not to the CS but to the masks: 
Perhaps the proposed misattribution process is independent of awareness of the CS but does require simultaneous presentation duration to exceed a certain minimum.
This may explain the lack of EC for masked CSs in Experiments 2 and 3. 
To explain why, in Experiment 4, the misattribution process operated on masked CSs when presented for 100 ms but not when presented for 30 ms, it may be assumed that the minimum presentation duration required for EC via misattribution lies between 30 and 100 ms. 
In effect, such a modified account would predict the absence of EC with subliminal presentation -- unless presentation duration can be extended to longer durations without simultaneously increasing awareness.
Masking CSs using continuous flash suppression allows for longer presentation durations and may be well suited to investigate this notion [for an overview of masking techniques see @breitmeyer_psychophysical_2015].
More generally speaking, however, whereas it is plausible that EC may *extend* to the mask stimuli, 
it is not easy to see why EC should be *restricted* to the masks and no longer operate on the CSs; 
EC effects would be of little practical relevance if they were so easily disrupted by the presence of other stimuli besides CS and US.
And if they were, dual-process models of attitude learning would still need to explain how one particular CS among several ones surrounding the US is singled out as the target for conditioning effects.

### EC for objectively unidentified CSs may be found with other masking techniques

We reliably and repeatedly failed to find EC effects for a specific presentation condition, namely pattern-masked CSs presented for 30ms.
Yet, EC for subliminal CSs may be obtained under other presentation conditions than those realized here:
Perhaps EC effects require a greater amount of basic perceptual processing of the CS.
Here we used a pattern mask to reliably disrupt processing of the CSs; this type of mask strongly interferes with basic processes in early areas of the visual processing stream and is therefore well-suited for the present purposes of disrupting conscious processing [@breitmeyer_psychophysical_2015].
Other masking techniques (e.g., metacontrast masking, object-substitution masking) interfere only at later processing stages; 
such techniques may be more likely to yield EC effects because they allow early visual processing to continue for a longer period while ensuring that a conscious representation of the masked stimulus cannot be formed [@breitmeyer_psychophysical_2015].

### EC may obtain for subjectively subliminal CSs

The present study used an objective threshold measure: Correct identification of a masked CS was taken to reflect evidence for perception of this CS above an objective threshold [@merikle_unconscious_1982].
However, it is possible that the identification task did not exclusively assess conscious processes, and that the above-chance identification performance obtained herein may be (in part or in whole) based on unconscious processes [@henley_unconscious_1982].
Put differently, the measure does not tell us whether participants in fact consciously perceived the correctly identified CSs -- that is, whether processing was not only above an objective but also above a subjective threshold [@cheesman_distinguishing_1986].
A correct response in the CS identification task may come about in different ways: 
Participants may have been fully aware of the CS; they may have perceived only one distinctive feature of the CS that allowed them to discriminate it from the other stimuli; or they may only have processed the stimulus unconsciously (i.e., without being aware of its identity) and may have selected the correct CS because it felt more familiar than the other options [e.g., @craik_recognition_2015]. 
In particular, some CSs may have exceeded an objective threshold (which enabled participants to correctly identify them based on a feeling of familiarity), but may have remained below the (higher) subjective threshold (i.e., participants did not consciously experience viewing this CS).
EC effects may obtain with such subjectively subliminal CS presentations (e.g., this may have been the case in the brightness-focus condition of the present Experiment 3; see Figure 6).
The present data do not allow us to address this issue of subjective awareness; future studies should investigate this possibility.

### EC may obtain for CSs presented below the objective identification threshold

The objective threshold/strategic (OTS) model of unconscious processes posits that unconscious effects can only be detected in the absence of conscious processes because conscious effects would override unconscious influences [@snodgrass_unconscious_2004].
More specifically, the OTS model postulates that unconscious perception is maximal at (or just below) the objective *detection* threshold, that is, when participants do not even notice the presence of the stimulus in question (this threshold is assumed to lie below the objective *identification* threshold investigated in the present research). 
In this case, conscious strategies to override unconscious processing are predicted to be absent.
Applied to the present research question, such a model would predict that unconscious EC might be able to operate in cases in which the presence of a CS is not even detected.
The model predicts that, when stimulus strength increases above the detection threshold, conscious overriding would initially reduce the effects of unconscious processes on EC (i.e., until the objective identification threshold is reached), before further increases in stimulus strength beyond the objective identification threshold would begin supporting EC effects due to conscious processes.
The latter predictions are in line with our findings. 
Yet, because the present study investigated the objective *identification* threshold we have no data regarding the objective detection threshold and therefore cannot test this model's critical prediction.
Thus, in principle, the lack of an EC effect for briefly presented CSs may be explained by conscious override.
Such an explanation can account for the present findings if unconscious EC effects can indeed be found at even lower levels of stimulus strength than those we realized here.
Despite the admittedly limited practical relevance of such unconscious effects that operate only in the complete absence of conscious processes [@reingold_unconscious_2004], it would certainly be worthwhile from a theoretical perspective to investigate the possiblity of EC effects for CSs at the objective detection threshold.


### EC may obtain with indirect evaluative measures or different CS-US pairing settings

The six experiments reported here relied on direct evaluative measures (i.e., evaluative ratings on Likert-type scales). 
We chose this measure for power-related reasons, as EC effects have proven much more sensitive to and robust on direct than indirect evaluative measures [see @hofmann_evaluative_2010].
One may argue, however, that indirect measures (e.g., based on the speed of responses) may be more sensitive for weak EC effects that may not be detected by direct evaluative ratings because they fall below a (conservative) criterion.
In addition, it may be argued that indirect evaluative measures such as the IAT or the affective priming task are more likely to reveal awareness-independent learning processes (e.g., because indirect measures tap associative representations, whereas explicit measures tap propositional ones).
Consistent with the latter view is a study by @rydell_two_2006 who found -- in the context of a subliminal US (not CS) study -- subliminal associative versus propositional learning procedures to selectively impact on indirect and direct evaluative measures, respectively. 
One should note, however, that both theoretical and empirical arguments speak against the latter suggestion. 
First, it is theoretically unclear why direct and indirect measures should dissociate as a function of propositional versus associative learning. 
For instance, outcomes of propositional learning may be automatized, leading to effects on indirect attitude measures, and attitudes claimed to be acquired via associative processes can be consciously accessed, allowing for conscious reports on direct evaluative measures [e.g., @gawronski_associativepropositional_2011; @gawronski_implicit_2014; @hahn_awareness_2014]. 
^[E.g., "people usually have experiential access to their affective gut reactions resulting from associative processes, and that they often rely upon these reactions in making propositional evaluative judgments" @gawronski_associativepropositional_2011, p.74.]
Second, dual-process proponents have repeatedly reported supportive evidence for awareness-independent learning that was obtained using direct evaluative measures [e.g., @hutter_dissociating_2012; @olson_implicit_2001]. 
And the opposite is also true, with awareness-depdendent learning effects obtained on indirect evaluative measures [e.g., @de_houwer_using_2006; @pleyers_aware_2007; @stahl_respective_2009].
Yet, as none of the current studies involved indirect evaluative measures, it remains an open empirical question for the future research whether subliminal CSs allow for successful EC when using indirect measures.

A similar comment applies to the pairing procedures used here. 
One proposed mechanism for awareness-independent evaluative learning about a novel CS is that participants implicitly misattribute the evaluative experience caused by the US to the (neutral) CS [@jones_implicit_2009].
For this mechanism to be operative, both CS and US stimuli should appear on screen simultaneously or in close temporal succession, so as to create a potential for confusion about the origin of the feeling; and participants should attend to the CS, so that the experience is likely to be misattributed to the CS [@jones_implicit_2009]. 
Attention to the US is assumed to be unnecessary; the valence associated with the US is presumably extracted and activated automatically [i.e., in the absence of attention to the US; @jones_implicit_2009].

@sweldens_evaluative_2010 proposed a similar mechanism for awareness-independent evaluative learning, namely the forming of direct (S-R) associations between CS and evaluative response. 
When such a direct link is formed, the CS can directly elicit the evaluative response without the intermediating role of the US representation. 
In addition to the simultaneous presentation of CS and US, the forming of direct S-R links is said to depend on another procedural factor: the pairing, of each CS, with multiple different USs of the same valence. 
This procedure has been implemented in most studies reporting evidence for awareness-independent evaluative learning.

In the current six studies, and in particular in Exp. 6, we met several requirements for allowing implicit misattribution or S-R learning. 
In all studies, CS-US presentations were temporally overlapping, and CSs were paired with USs of the same valence but different identities. 
In addition, in all studies except Expts. 2 and 3, attention during learning was focused on CS identification -- as in the surveillance paradigm -- or on non-evaluative features of the US (i.e., the brightness focus).

It is however possible that even stronger requirements would need to be met for implicit misattribution to operate on subliminal CSs. 
Specifically, subliminal CSs may lead to successful EC when the following conditions are *jointly* met: 
(1) USs of weak or moderate intensity (as more intense USs prevent implicit misattribution), 
(2) USs of low perceptual salience (as more salient USs prevent implicit misattribution), 
(3) identical CS-US onsets (as presenting the USs before the CSs may increase US salience and so prevent implicit misattribution), 
(4) incidental learning (as explicit learning prevents implicit misattribution), 
(5) use of indirect evaluative measures (see discussion above).

Three comments are in order here. 
First, even though our studies may not be optimal for eliciting implicit misattribution effects, they are arguably more adequate than the modalities used in previous studies that reported significant EC effects with 'subliminal' CSs -- yet involved, for instance, sequential and non-overlapping CS-US pairings. 
Second, meeting collectively the above conditions may, indeed, be conducive of EC with subliminal CSs. 
Yet, these conditions would then appear to be so constraining that they would inevitably lead to question the theoretical and practical relevance of dual-process approaches to attitude learning. 
Third, and perhaps even more important, it is not even clear whether the implicit misattribution account would predict any EC effect under subliminal conditions.
As a matter of fact, @jones_implicit_2009 argued and showed that EC based on implicit misattribution requires that participants shift their gaze between the CS and the US so as to visually connecting the two -- this was supported using both an eye-tracking procedure and a flashing procedure that drew participants' attention to the CS-US co-occurrence.
It is unclear how the gaze-shifting requirement may be implemented in a subliminal setting.

Given these open questions, our findings do not conclusively rule out the possibility of awareness-independent evaluative learning.
Yet, they point out clear limitations to the scope of such effects that will guide theorizing and help build more precise models of evaluative learning.

## Practical and theoretical implications

General claims that EC is independent of awareness may be specified in different ways. 
Using a framework of conscious perception by @dehaene_conscious_2006, we separated attention and stimulus strength as important contributions to awareness in EC. 
We manipulated stimulus strength and showed that EC does not occur for brief and masked presentation of CSs. 
Thus, in terms of this framework, EC does not seem to be possible under subliminal (but attended) conditions. 
This finding suggests that EC should also be absent under subliminal and unattended conditions because stimulus-related activation, which may have been increased somewhat via top-down attention in the present study, should be even weaker without such top-down attention. 
The present findings suggest that EC is dependent on awareness in the sense that it cannot operate on perceptually unidentifiable CS stimuli.

The present findings do not deny the possibility that EC may occur under supraliminal but unattended conditions, a condition termed ‘preconscious’ in the above framework. 
Preconscious processing is characterized by strong stimulus-related activation (that may, e.g., trigger priming effects), but this activation is limited to sensorimotor areas and does not extend to frontal areas involved in attention and conscious perception; as a consequence, the stimulus is not consciously represented or reportable [@dehaene_conscious_2006]. 
However, because preconscious stimuli are thought to become conscious once attention is directed towards them, it is difficult to assess the level of consciousness, using on-line self-report measures of awareness as in the present approach, without interfering with the preconscious state of the stimuli. 

The possibility of preconscious EC is compatible with recent findings claiming to show EC effects in the absence of awareness.
Most prominently, the work in the surveillance paradigm has repeatedly demonstrated EC effects under incidental learning conditions [e.g., @olson_implicit_2001; @stahl_incidental_2016]. 
In that work, both CSs and USs are clearly identifiable, and participants were required to attend to the individual stimuli (in order to detect the presence of a specific target picture) but not to CS-US pairs or co-occurrence relations. 
However, participants later reported being unaware of the co-occurrence relations or contingencies between the stimuli (i.e., the fact that one CS was paired only with positive USs, and another CS paired only with negative USs). 
EC effects in this paradigm may reflect preconscious learning because, while participants processed clearly identifiable CS and US stimuli, they apparently remained unaware of the critical CS-US co-occurrence information. 

For practical purposes, the meta-analysis of the present findings suggests that -- although small -- EC effects are reliable in incidental learning situations, even when people are neither instructed to attend to CS-US co-occurrences nor to attend to the valence of the US. 
In such situations, people are often busy performing a different and unrelated task, and their intention is not one of learning about the unfamiliar CS stimuli; nevertheless, incidental evaluative learning processes may still operate.
If, under such incidental and largely unattended conditions, participants' evaluations of these novel CSs are susceptible to modulation by the valence of the paired USs, these US valence effects can be characterized as unintentional (i.e., occurring in the absence of an explicit intention to learn).
^[The present data do not speak to the question whether such effects are also uncontrollable in the sense that participants are unable to resist such influences if they were made aware of their possible existence.]
Such incidental and unintentional effects of US valence are thought to be at the basis of some advertising effects [e.g., @biegler_feeling_2016].
While advertising may exert positive influences for producers, consumers benefit if they are aware of its potential influence on their evaluations and decisions.
For instance, advertising can have serious negative side-effects when advertisements for prescription pharmaceuticals lead to an unwarranted increase in the use of those drugs [@biegler_ban_2013].
It would be of great practical relevance to investigate, for instance, whether warnings about the possible unintended and incidental influences of advertisements are effective in limiting such undesired side-effects.
The present study suggests that practitioners and applied researchers should focus their concerns on EC for supraliminal stimuli under incidental learning conditions.