Stronger Suboptimal than Optimal Affective
Priming by Faces:
A meta-analysis
Wenyu Nie
Student number: 11120053
First Internship Research Master Brain and Cognitive Sciences
Track: Cognitive Neuroscience
Amsterdam Brain and Cognition Center
University of Amsterdam
Supervisor: Dr. R.H. Phaf
Second assessor: Dr.
R. Grasman
Credits: 26 EC
Abstract
Affective priming refers to the phenomenon that the presentation of a valenced prime stimulus influences the subjects' affective decision of a subsequent target stimulus. Murphy and Zajonc (1993) found that affective priming, specifically by valenced faces, was larger in
suboptimal/nonconscious prime presentation conditions than in optimal/conscious conditions. This finding highlights the automatic nature of emotion processing and its evolutionary advantage. In addition, it points to a qualitative difference between conscious and nonconscious processing. The suboptimal advantage of affective priming is, however, not always replicated in later studies of facial affective priming. This may result from publication bias or potential hidden moderator variables. We present a meta-analysis of 29 facial affective-priming studies with suboptimal, optimal, or both conditions. All 49 effect sizes were calculated from the means and standard deviations in these studies (sample sizes combined N= 2047) to examine the effects of affective priming in the optimal and suboptimal conditions, separately. Method to obtain suboptimality, negative prime type, target type, processing strategy, and task served as moderator variables for the meta-analysis. Results showed a small-to-medium (g=0.2850) effect in the suboptimal condition and a small (g=0.1781) effect in the optimal condition after correction for publication bias. Backward masking produced larger effect than brief presentation. Studies using sad faces as
negative primes produced larger effects than those using other negative primes. In liking rating task, neutral faces, compared to neutral ideographs, elicited larger effects. Surprisingly, the moderator task was confounded with the suboptimal-optimal manipulation. When prime presentation is conditioned on the moderator task, the suboptimal advantage disappears. Although no study explicitly manipulated processing strategy, it was tentatively concluded that processing strategy could better explain these findings than the suboptimal-optimal manipulation of conscious processing.
Stronger Suboptimal than Optimal Affective Priming by Faces: A meta-analysis
About half a century ago, a public upset arose from the findings reported by the market researcher James Vicary (1956). He claimed to have found that movie audiences could be surreptitiously controlled by invisible (i.e., nonconscious) messages encouraging them to “Drink Coca-Cola” and “Eat popcorn”. Laws were even enacted in some states to ban subliminal advertising (Goodkin & Phillips, 1980). More than three decades later, television commercials, magazine ads, and
bookstores promoted subliminal tapes that promised to induce dramatic improvements in mental and psychological health. These devices were widely advertised as producing many desirable effects, including weight loss, memory enhancement, and improvement of sexual function. The subliminal advertising in the theaters turned out to be a hoax and the later subliminal self-help audiotapes were shown to be totally useless with regard to therapeutic benefits (Greenwald, Spangenberg, Praktkanis, & Eskenazi, 1991). Despite these false initial attempts in the applied domain, there is a considerable amount of evidence in laboratory research that subliminal stimuli can have appreciable effects. However, these findings were frequently confronted with failed replications and criticisms on measurement validity (Cesario, 2014). This meta-analysis attempts to find out whether these effects in the field of emotion really ‘stick’, and are not a result of
confounding factors, sloppy measurements, or publication bias.
One prominent candidate for nonconscious operation is emotional processing. Because affective reactions to some types of emotional stimuli seem to be evolutionarily prepared (Öhman, 1986; LeDoux, 1996), they may be more basic than, and even precede, cognition (i.e., the Affective Primacy hypothesis; Zajonc, 1980). This evolutionary privileged status may also make them more liable to nonconscious (i.e., subliminal or suboptimal) processing than other types of processing. Particularly for emotional facial expressions, there is strong neuro-imaging (e.g., Lapate et al., 2016; Vuilleumier, Armony, Driver, & Dolan, 2003; Whalen, Rauch, Etcoff, McInerney, Lee, & Jenike, 1998), electrophysiological (e.g., Liddell et al., 2005; Pourtois, Spinelli, Seeck, &
may be subject to direct, nonconscious processing. In a seminal study, Murphy and Zajonc (1993) illustrated the independence of affect from cognition by showing that the affective priming by valenced faces was stronger in suboptimal (i.e., with reduced levels of conscious processing of the face primes) than in optimal (i.e., fully conscious processing of the faces) conditions.
Priming refers to the phenomenon that the presentation of one stimulus (i.e., the prime) alters perception of a second stimulus, the target. Murphy and Zajonc investigated affective priming on the evaluation of meaningless Chinese ideographs preceded by emotional facial expressions. Cognitive priming was implemented by non-affective primes (i.e., size, symmetry, gender). For affective priming, participants were instructed to rate their subjective preference for the ideographs (Experiment 1), or to evaluate the affective meaning of the ideographs (Experiment 2) with the priming of affective faces. For cognitive priming, they were instructed to evaluate the physical features (Experiment 4, symmetry), or non-affective meaning (Experiment 3, size; Experiment 5, gender) of the ideographs with non-affective shapes and faces. They then juxtaposed affective priming and cognitive priming across brief (suboptimal) and longer (optimal) prime-exposure durations. If affective reactions, as is suggested by affective primacy, are more immediate and less under voluntary control, it could be expected that emotion-laden stimuli presented in suboptimal conditions would influence the subsequent judgment or evaluation more strongly than in optimal conditions. This hypothesis was clearly supported by their results that affective primes produced shifts of judgments on the evaluation of meaningless stimuli at suboptimal exposures but not at optimal exposures.
These findings are also important from the perspective of conscious-nonconscious processing (cf, Merikle, 1992). There has been a long-standing debate about whether conscious processing and nonconscious processing are qualitatively different or lie at different points of a continuum (e.g., Sergent & Dehaene, 2004). It is believed by some researchers that nonconscious processing purely represents a dilution of conscious processing so that the magnitude of any effect should always be reduced by this dilution (e.g., Desender & van Bussche, 2012). The findings of
Murphy and Zajonc clearly contradict this view, because they showed larger suboptimal (i.e., with reduced conscious processing) than optimal (i.e., fully conscious) effects. Deciding whether the affective priming pattern obtained by Murphy and Zajonc really sticks is thus highly relevant for the question whether nonconscious facial affective processing qualitatively differs from conscious affective processing, and whether conscious vs. nonconscious processing has any meaning.
To demonstrate unconscious psychological processing, researchers usually contrast a direct measure, checking the awareness of a relevant stimulus, with an indirect measure, showing that the stimulus has its influence on performance and therefore has been processed. Most studies in the field attempted to elicit indirect psychological effects with the direct awareness entirely suppressed. Merikle (1992) has argued that this classical indirect-without-direct effect convention hinges on the complete exclusion of all conscious processing in the direct task. Moreover, accepting the null hypothesis of direct task performance on the basis of nonsignificance can be a very hazardous exercise (Hartgerink et al., 2016; Vadillo et al. 2015). Because of the spurious assumptions in the simple dissociation studies, Merikle suggested that a qualitative dissociation (i.e., reverse effects under conscious and less conscious condition) is better suited to show that consciousness really matters. Schmidt and Vorberg (2006) further discussed three possible types of dissociation between the indirect and direct measures. According to their classification, simple dissociations between the direct effect and the indirect effect occur when the indirect effect is larger than zero and the direct effect is at chance level, which is the common approach in the field; sensitivity dissociations occur when the indirect effect is larger than the direct effect; double dissociations occur when the direct and indirect effects are going opposite direction. Double dissociations require much less
assumptions than simple or sensitivity dissociations and it is already sufficient to indicate
qualitative difference between conscious and unconscious processing. Schmidt & Vorberg (2006) depicted these three types of dissociations by plotting the effect sizes of the indirect task against the direct effect sizes of the direct task (D-I plots).
Some later studies seemed to corroborate the seminal findings of Murphy and Zajonc. For example, Rotteveel, de Groot, Geutskens, and Phaf (2001) measured affective priming with electromyography (EMG) measures. Facial EMG is a very sensitive and valence-specific measure of affect. Facial EMG of the musculus zygomaticus major for smiling and musculus corrugator supercilii for frowning are two useful indices for the measurement of valenced states. With the facial EMG recorded after the onset of the primes, the biased ratings were more solidly traced back to the influence of the affected facial primes. Furthermore, many studies employed the suboptimal affective priming paradigm for other research purposes, and obtained supportive results in the process. These include studies for social judgments (Hooker, Tully, Verosky, Fisher, Holland, & Vinogradov, 2011), emotion-related impairments in psychopathology (Suslow, Roestel, & Arolt, 2003), empathy (Iacono, Ellenbogen, Wilson, Desormeau, & Nijjar, 2015) and consumer behavior (Winkelman, 2005).
The suboptimal advantage was not always reproduced. Sometimes, the affective priming effect did not even show up in the suboptimal condition. For instance, in an attempt to replicate the Murphy and Zajonc findings with almost exactly the same experimental settings, the opposite phenomenon was found. The affective primes only produced significant shifts in preference judgments of novel stimuli at longer exposure durations. At suboptimal exposures the judgments were not influenced by the affective primes at all (Kemps, Erauw, & Vandierendonck, 1996). In addition, informal contacts with other researchers also suggested that many non-replications may have been suppressed due to the nonsignificance of the results. It is hard to tell how many studies have been conducted, and how many turned out nonsignificant and were never reported. In fact, this selective publication of positive results, called publication bias, has been a big problem in the whole field of psychology (Rosenthal, 1979). A recent initiative to replicate a large number of classical psychological findings even revealed that a majority could not be reproduced (Open Science Collaboration, 2015). Even without any questionable research practices or bias, the majority of published, significant findings was calculated to be false by Ioannidis (2005). Not only the false
positive rate, but also the false negative rate (i.e., rejecting nonsignificant but “true” effects) may be excessively, and shockingly, high. Hartgerink, van Assen, and Wicherts (2015) in their calculations revealed the possibility that some of the nonreplications in the open science collaboration was due to such false negatives.
The difficulty in replicating Murphy and Zajonc may also stem from some hitherto unrecognized, but crucial, hidden variable (Klein, 2014). Klein noted that the hidden moderators invoked by nonreplicated researchers may sometimes appear trivial (e.g., testing in cubicles), and unrelated to theory (see also Yong, 2012), and much recent research in affective priming and related areas (e.g. mere exposure) suggested more meaningful factors. Likely candidates for such variables in the affective priming paradigm are processing strategy (Whittlesea & Price, 2001), method to obtain suboptimality (e.g., Almeida, Pajtas, Mahon, Nakayama, & Caramazza, 2013), negative prime type (e.g., Marsh, & Ambady, 2007), task, and target type, which is nested under task. (Wedell, Parducci, & Geiselman, 1987). A more extensive discussion of these moderator variables can be found below.
Method to obtain suboptimality
Non-replications may be caused by different methods to obtain suboptimality. This was vividly shown by a study (Almeida, Pajtas, Mahon, Nakayama, & Caramazza, 2013) that has put continuous flash suppression (CFS) and backward masking in juxtaposition. In CFS, one eye is presented with a static stimulus and the other eye is presented with a series of rapidly flowing (i.e., flashing) stimuli, which suppress the static stimulus. In backward masking (BM), a visual stimulus (i.e., the mask) is presented immediately after a brief target stimulus which leads to a failure to consciously perceive the target. In spite of the equivalent phenomenological suppression of the visual primes in CFS and BM, the results were slightly different with these two masking techniques. Although suboptimal affective priming was obtained for both angry and happy invisible facial expressions with backward masking, it showed up only for angry facial expressions with continuous
flash suppression. If this holds true, it can be expected that the effect size for affective priming with CFS manipulation will be smaller than that with backward masking manipulation, because with CFS, the effect is only obtained one way from negative primes, whereas with backward masking, the effect is elicited from both positive and negative directions.
It was argued by Merikle and Joordens (1997) that divided attention could also produce suboptimal conditions. This claim was empirically supported by the affective priming study of Rotteveel and Phaf (2004). These authors manipulated the working memory load by operating a secondary task of retaining digit-letter strings. The stronger affective priming in this suboptimal condition (divided attention), which is not even claimed to be fully nonconscious, than in an optimal condition (focused attention) revealed the reverse effect strength of affection on the continuum of consciousness, compared to non-affective cognition. Merikle (1992) has argued that no research on nonconscious processing whatsoever can guarantee to exhaustively exclude all conscious processing, and therefore, no single method for obtaining suboptimality may able to achieve fully nonconscious processing (Merikle, 1992). Nonsignificance in the tests for conscious processing seems generally due to underpowered statistical tests and the occurrence of false
negatives in many direct tasks (Vadillo, Konstantinidis, & Shanks, 2015; see also Hartgerink et al., 2016). There can thus be systematic differences between the methods in achieving suboptimality, which may not be reflected in the results of the direct task. This will show up in larger effect sizes for the most efficient methods, if indeed suboptimal are larger than optimal priming effects. If however optimal priming is larger than suboptimal priming, the most efficient methods will produce the smallest effect sizes. Therefore, this meta-analysis motivates the moderator variable, method to obtain suboptimality, and expects it to have four levels: brief presentation, masking, CFS, and divided attention.
Whittlesea and Price (2001) observed in their study of mere exposure that the dissociation of preference and recognition performance might be caused by a confound of analytic and nonanalytic processing strategies. A non-analytic strategy corresponds to treating stimuli as a whole and looking only at the global features. An analytic strategy, on the other hand, entails a detailed analysis of the stimuli and paying attention to small-scale, local, properties. They argued that preference judgments implicitly elicit non-analytic strategies, whereas recognition more often involves analytic
processing, because for the latter similar target stimuli have to be discriminated, whereas the former rely on global impressions. In support of their reasoning, they were able to control the emergence and disappearance of both preference and recognition performance by instructing non-analytic and analytic processing strategies, respectively.
The reasoning of Whittlesea and Price (2001) may also extend to affective priming studies. Optimal presentation of the facial primes may induce analytic processing that can lead to the affective separation of primes and targets and a lack of affective transfer from prime to target. A non-analytic strategy can be implicitly elicited by suboptimal presentations, because it is generally not possible to look into the details of the prime. This type of processing may facilitate the affect spilling over from prime to target. Results of this effect with word primes were obtained by Alexopoulos and his colleagues (2012, 2017), when explicitly manipulating processing strategies. Therefore, to see if the influence is also there for facial prime cases, this meta-analysis motivates the moderator variable, processing strategy, and expects it to have two levels: analytic processing and non-analytic processing.
Negative prime type
The most commonly used negative prime faces in affective priming studies are sad, fearful, or angry expressions. Also sometimes the exact nature of the negative prime faces is left unspecified by the researchers, which may suggest that mixtures of these expressions were included. The specific type of negative facial expression may also be able to steer the direction and degree of
affective priming. Fearful, angry, and sad (Rotteveel & Phaf, 2004; Carver & Harmon-Jones, 2009; Krieglmeyer & Deutsch, 2013) expressions can all have ambiguous emotional meanings. Distress cues, such as fearful or sad faces, can be associated with prosocial responding (i.e., approach for soothing the fear; Marsh & Ambady, 2007), and thus may also elicit some positive affect. For instance, Marsh, Ambady, and Kleck (2005) found that perceivers’ dominant response to a fearful facial expression is approach, due to its shared features with infantile faces (Hammer & Marsh, 2015). The negative affect associated with angry faces may be stronger than with fearful faces, but may evoke both avoidance-related fear and approach-related anger. Even in the interpretation of angry faces thus some ambiguity remains, which may reduce their suitability as negative prime. In addition, it has recently been argued that the perception of facial expressions does not fall in distinct categories, but relies heavily on contextual factors (Barrett, Mesquita, & Gendron, 2011). These confusions make it hard to purely infer which negative emotion will be the best candidate. This meta-analysis therefore includes negative prime type as a moderator variable, and expects it to have four levels: sad faces, fearful faces, angry faces, or unspecified.
Task
Different types of task used for eliciting affective priming can in some way boost or diminish the effect. Liking rating task was the task originally used by Murphy and Zajonc (1993) in their seminal study for measuring the effect of the affective primes on targets. In this type of task, targets would be set to be neutral and any valence attached to it would be totally out of the affective priming effects. However, some other studies (cf. Fazio, 2001) took use of the congruency effect and
demonstrated that the time needed to evaluate target stimuli as positive or negative decreased if they were preceded by a similarly valenced prime stimuli. In this type of task, targets would be set to valenced and the effects of the primes would be measured through its facilitation or interference on the reaction time for the valence decision of the target stimuli. Moreover, some other studies (cf. Rotteveel et al., 2001), as is mentioned above, measured facial electro-myography (EMG) in
addition to affective ratings. Considering the sensitivity of the EMG measurement, it might be expected the effect size would be larger for EMG cases, rather than liking rating cases or valence decision cases. This meta-analysis includes task as a moderator variable, and expects it to have three levels: liking rating task, valence decision task, and EMG measurement task.
Target type
Both neutral faces and meaningless stimuli serve as targets in liking rating experiments. It has been found that the evaluation of target faces depends on the affective nature of the preceding facial stimuli (Wedell, Parducci, & Geiselman, 1987). Neutral faces would be evaluated as more positive after negative facial expressions and more negative after positive facial expressions. This successive contrast effect could complicate facial affective priming by counteracting the priming effects. Therefore, it can be expected that smaller effect size will occur in those studies using neutral faces as targets for ratings instead of meaningless neutral shapes. Yet another problem of neutral faces is their inherent negative meaning. Lee, Kang, Park, Kim, and An (2008), using the Extrinsic
Affective Simon task (EAST), found that neutral faces were evaluated similarly as negative faces. The study suggested that prototypical ‘neutral’ faces are actually evaluated as negative in many circumstances, which further suggests the inclusion of neutral faces as centralized unbiased targets might introduce a confound in the experimental design. Therefore, the meta-analysis motivated target type as a secondary moderator variable nested under task. For liking rating task and EMG measurements, three levels, neutral faces, nonsense words, neutral ideographs; for valence decision task, two levels, emotional faces, emotional words (emotional musical pieces added post hoc).
In sum, the empirical evidence in favor of and against stronger suboptimal-than-optimal affective priming by faces hangs in the balance, and the opposing views may point to one or more crucial moderating factors. The only way to embrace the evidence from the two sides and to
consider the potential moderators is to perform a systematic meta-analysis. We first investigated the hypothesis that affective priming of faces is stronger in suboptimal than in optimal conditions, and
then analyzed this further into four different moderator variables. It should also be revealed in the process whether, and to what degree, the field suffers from publication bias.
Method
Search procedure
The literature search was performed according to PRISMA (Moher, Liberati, Tetzlaff, & Altman, 2009) and the search process with the resulting number of studies was specified in a flowchart. Four databases (ISI Web of Science, PsycINFO, PubMed, and Google Scholar) were searched for
affective priming papers. The search was performed using the search string “(affective priming OR evaluative priming) AND face,” with OR and AND representing Boolean operators in ISI Web of Science, PsycINFO, and PubMed. In Google Scholar, cite reference searches were conducted in order to obtain studies that referred to the seminal article of Murphy and Zajonc (1993). Duplicates resulting from the different searches will be screened, and theoretical papers, not involving new empirical research, were removed.
Inclusion and exclusion criteria
Studies were included and excluded according to the criteria below, applied in this order:
1) Only published English-language studies, appearing mostly in peer-reviewed journals, were included. Studies that are unpublished, retracted or under strong suspicion of fraud or questionable research practices (i.e., being formally investigated), were excluded.
2) Only studies using a valence decision task, like-rating task or EMG measurement to investigate the short-term priming effects on a trial-by-trial basis were included. Studies using other tasks (e.g. pronunciation task, moral-decision task) or measurement (e.g. event-related Potential), or investigating long-term priming effects were excluded.
3) Only studies with healthy participants, also healthy controls in clinical studies, were included. Studies involving only clinical patients selected for some specific trait or characteristic were excluded.
If the selection is made a-posteriori (e.g., with a median split) on an unselected sample, whole-group results were included. Also studies involving state (e.g., mood) inductions were excluded. Control groups in these studies, without such an induction, were included.
4) Only studies with face primes were included; studies with other types of primes (e.g. non-face pictures, words) were excluded.
5) Only studies with both positive face primes (i.e., happy) and negative face primes (i.e., angry, fearful, or sad), were included. Studies with only negative or positive primes were excluded.
6) Only studies providing means and standard deviations were included. Studies providing only statistical test parameters (e.g., t-value or F-value) were excluded. Attempts were made to contact the authors of the latter studies for the means and standard deviations. If they did not react or could not provide these, the study was excluded.
Only studies reporting results for different levels of a moderator variable separately were included at that level. Studies mixing up the results for different levels of a moderator variable were excluded from the level of that moderator variable on in the specific moderator analysis.
7) Next to the indirect (i.e., priming) measure, all studies should include, some indication of the level of conscious processing. When the authors, for whatever reason, indicated that the results were subliminal, suboptimal, unconscious, nonconscious, or without awareness, this effect was classified in the suboptimal level. All other studies, including those without indication of the level of conscious processing, were classified in the optimal level. If the studies contained a direct test, its effect sizes would be included in the D-I graph.
Moderator variables
Hypothesized categorical moderators are (1) method to obtain suboptimality: brief presentation, divided attention, masking, flash suppression, spatial frequency; (2) processing strategy: analytic, non-analytic; (3) negative prime type: sad, fearful, angry, unspecified; (4)task: liking rating, valence decision, EMG measurement; (5) target type, nested in task: for liking rating task and EMG measurements, three levels, neutral faces, nonsense words, neutral ideographs; for valence decision
task, two levels, emotional faces, emotional words (emotional musical pieces added post hoc). The statistic Qm served as an omnibus test for differences between levels, Qe as a test for residual
heterogeneity. In the result, some levels were missing because of there were none cases falling in them.
Effect size calculation
Due to the slight bias of Cohen’s d in small samples (Hedges & Olkin, 1985), effect sizes in this meta-analysis were computed in terms of Hedges’ g. (See Equation 1).
pooled negative positive
S
M
M
g
(1) Hedges’ g is different from Cohen’s d in that it uses n-1 instead of n for each sample variance when pooling the sample variance. It therefore provides a better estimate, especially for small sample sizes (Hedges & Olkin, 1985). In the context of this meta-analysis, as is shown in the Equation 1, it is calculated by dividing the difference between the mean performance in the positive and negative conditions by the pooled standard deviation. The positive condition refers to situations in which participants are asked to perform (liking rating, valence decision) after being primed by positive facial expressions. The negative condition refers to situations in which subjects are asked toperform the tasks after being primed by negative facial expressions. Performance is operationalized through the liking index, reaction time or muscle activity amplitude according to different
dependent variable measurements in different tasks. It should be noted that for some studies using the valence decision task to investigate affective priming the comparisons were made between congruencies and conflicts, but the mean reaction times (RT) for positive affective priming and negative affective priming were not reported separately. The effect size in these cases was calculated through another equation from the mean reaction times in congruent and incongruent conditions, which corresponds to Equation 1(See Equation 2).
pooled congruent t incongruen
S
M
M
g
(2)To clearly visualize the potential dissociation of conscious and nonconscious processing, the effect sizes of the direct task were also collected, and the effect sizes of the indirect task were plotted against the effect sizes of the direct task (Schmidt & Vorberg, 2006). The effect sizes of direct effects were collected from the awareness check, either from the mean sensitivity d’ and its standard deviation or from the p values of the tests.
Repeated measures correlations
Because all studies had repeated measures designs, the correlations between the repeated observations within subjects were needed for the computation of g (Borenstein et al. 2009). However, it is not common for researchers to report this statistic in their studies. In this meta-analysis, it was estimated from the t value in paired t-tests (see Equation 3) or the F value in repeated measures ANOVAs (see Equation 4).
2 2 2
2
2
t
n
g
t
r
(3)F
n
g
F
r
2
2
2
(4)For the studies that did not report test statistics, the average of all available correlations weighted by individual sample sizes was imputed as the correlation for that individual study.
Data analysis
The meta-analysis was performed with the package metafor (Viechtbauer, 2010) embedded in the software R (version 3.2.5-3.3.2) (R Development Core Team, 2010). Analyses of the effect sizes were performed with the Random-effects model, in order to catch the potential heterogeneity of the effect sizes. Moderator variables had been motivated and defined a priori and analyzed in the mixed
effects model (Borenstein et al., 2009). Separate analyses were performed for optimal and suboptimal conditions. The proportion of systematic unexplained variance (τ2
) was estimated using the restricted maximum-likelihood estimator (REML) (Viechtbauer, 2010). Cochran’s Q-test
(Hedges and Olkin, 1985) were used to test the heterogeneity of the effect sizes. Significance of the test would indicate heterogeneity of effect sizes from different studies. Influence analysis (i.e., the exclusion of single studies) was performed to identify outlier studies (Cook & Weisberg, 1982).
Publication bias
Funnel plots and p-curve methods were used as the tool to examine publication bias. In funnel plots, a measure of accuracy (in our case, sample size) was plotted against the effect sizes. Without
publication bias, the dots representing each study would scatter symmetrically around the effect sizes of the most accurate (the ones with largest sample sizes) studies like a triangle. When non-significant results remain unpublished, an asymmetrical funnel plot should appear. A regression test for funnel plot asymmetry was conducted within subset studies of different optimality conditions, in order to check the occurrence of publication bias in these subsets (Egger & Smith, 1997). To correct for the possible publication bias, the trim-and-fill method was applied (Duval & Tweedie, 2000) to estimate the number of missing studies and adjust the overall effect sizes.
The p-curve method assumes that publication bias does not apply to the effect sizes but to the significance levels. It detects deviant significance patterns and thus corrects for publication bias. The p-curve is a distribution of statistically significant p values for a set of relevant studies
(Simonsohn, Nelson, & Simmons, 2014a). It can be expected that true effects will generate a left-skewed p-curve containing more high than low significant p values. The p-curve method corrects for publication bias without resorting to non-significant studies that may not have been published at all (Simonsohn, Nelson, & Simmons, 2014b). This was especially advantageous in this domain, for significant studies vastly outnumbered non-significant studies in the field of affective priming. The
frequencies of different values of p were drawn in histograms and compared to the prototypical frequency distribution in the studies of Simonsohn, Nelson, and Simmons (2014ab).
Results
The literature search was done according to the PRISMA procedure (Moher, Liberati, Tetzlaff, & Altman, 2009) and the search process was visualized in a flowchart (see Figure 1). The search strategy, as is described in the method section, produced 210 study cases in google scholar, 274 in web of science, 127 in PsycINFO, and 138 in Pubmed. After the online screening of the irrelevant articles and duplicates, there were 198 cases kept to be screened in full text. Full-text screening excluded 132 studies for not satisfying the inclusion and exclusion criteria. 39 of the remaining 66 cases lacked crucial statistical values and authors were contacted. Two of them replied and they were thanked in the acknowledgment. The other 37 cases were excluded. Eventually, 29 studies (49 effect sizes) were included in the meta-analysis.
Figure 1. The flowchart of screening according to PRISMA procedure
In the forty-nine effect sizes, 18 were effect sizes in the optimal condition and 31 were effect sizes in the suboptimal condition. Thirty studies employed the liking rating task and the other 19 the valence decision task. In terms of which type of negative stimuli were used as prime, the numbers of studies using sad, fearful, angry and unspecified stimuli were 19, 13, 9, and 8 respectively. No studies distinguished processing strategy of participants and the moderator was dropped in the further analysis. No studies using EMG measurement survived the inclusion and exclusion criteria either, this level of moderator task was also dropped in the further analysis.
Table 1. Counts of effect sizes in different levels of each variable
Moderator Level k
Optimality optimal 18
suboptimal 31
Task liking rating 31
valence decision 18
Negative prime sad 19
fearful 13
angry 9
unspecified 8
Method back masking 16
brief presentation 11
CFS 2
Divided attention 1
Target Liking rating Neutral faces 8
Nonsense words 1
Neutral ideographs 22
Valence decision Emotional faces 10
Emotional words 7
Emotional music pieces 1
Overall effect
The effect sizes (k=49) ranged from -0.34 to 2.51, revealing an overall small-to medium effect size (g=0.358; 95% CI [0.242, 0.473]: p < 0.001). Most of the effect sizes were in the positive direction
(k = 42). There was a large and clear amount of heterogeneity in the effect sizes (Q = 333.69, df = 48, p < 0.0001). It is estimated to be 𝜏2=0.143, 95% CI [0.122, 0.3845]. Taking optimality as the
moderator variable in a mixed effect model did not reduce the amount of heterogeneity much (Q= 329.211, df = 47, p < 0.0001). There was still a considerable amount of residual
heterogeneity 𝜏2=0.148, 95% CI [0.122, 0.384]. Further analyses were performed for the separate optimality levels.
Optimal condition
The effect sizes (k=18) range from -0.34 to 0.71. Testing with the random effects model yielded an overall effect size of g = 0.270 (p < 0.0001; 95% CI [0.154, 0.385]). Most of the effect sizes were larger than zero as expected (k=16). (See Figure 1). The heterogeneity was estimated to be 𝜏2=0.042; 95% CI [0.0154, 0.1352]. There was a considerable amount of heterogeneity of effect
sizes in the optimal condition (Q=62.12, df = 17, p < 0.0001). Outlier and influence analysis found one outlier study, g = -0.34 (studentized residual = -2.281; Kim, Lee, Ha, Kim, An, Ha, Cho, Hyun-Sang, 2011, the optimal condition). Exclusion of this outlier increased the overall effect size to g = 0.29 (p < 0.001; 95% CI [0.18, 0.40]). The heterogeneity was reduced, but still remained at a
considerable level (Q = 54.9968, df = 16; p < 0.0001). All following analyses were conducted under exclusion of this outlier.
Moderator analysis of optimal condition
Due to the fact that the outlier was the only masked case in the optimal condition, there were only two moderator variables (task, negative prime type) included in a mixed effects model. The estimated amount of residual heterogeneity 𝜏2=0.0145; 95% CI [0.0011, 0.0900]. The reduced 𝜏2
suggested that 55% of the heterogeneity was explained by the moderator variables tested here (𝑄𝑚= 8.1576, 𝑑𝑓 = 2, 𝑝 = 0.0169). However, the test for residual heterogeneity showed that
there was still a considerable amount of heterogeneity (𝑄𝑒 = 28.001, 𝑑𝑓 = 14, 𝑝 = 0.0142). Details of the moderator analysis can be found in Table 2. The moderator variable task could explain some of the heterogeneity (𝑄𝑚 = 3.5613, 𝑑𝑓 = 1, 𝑝 = 0.0591). The average effect size
was g=0.4290 for the liking rating task (k= 6, 95% CI [0.2593, 0.5987], p<0.0001), and g = 0.2263 for the valence decision task (k=11, 95% CI [0.1058, 0.3467], p = 0.0002). The moderator task explained 27.61% of the heterogeneity in the effect sizes (𝜏2 = 0.0234). The test for residual heterogeneity showed that still a lot of heterogeneity remained (𝑄𝑒 = 36.4891, 𝑑𝑓 = 15, 𝑝 < 0.0015).
Table 2. Results of moderator analyses of the optimal condition
Moderator Level k Estimate [95% CI] p
Task Liking rating 6 0.4290 [0.2593, 0.5987] <0.0001
Valence decision
11 0.2263 [0.1058, 0.3467] 0.0002
Negative prime sad 7 0.4153 [0.2580, 0.5726] 0.0001
fearful 4 0.2466 [0.2580, 0.5726] 0.0253
angry 3 0.2141 [-0.0104, 0.4385] 0.0616
The affective priming by sad faces reached medium levels, whereas the other negative faces only showed small effects. The task of the moderator variable negative prime type could explain some of the heterogeneity (𝑄𝑚= 6.5176, 𝑑𝑓 = 1, 𝑝 = 0.0107). The moderator task explained
53.54% of the heterogeneity in the effect sizes (𝜏2 = 0.0150). Still a considerable amount of heterogeneity remained (𝑄𝑒 = 30.9376, 𝑑𝑓 = 15, 𝑝 = 0.009).
Due to the different nature of the liking rating task and the valence decision task, the moderator variable target type was analyzed separately for the two tasks. For the liking rating task cases, the moderator target type could not be analyzed, for there was only one level of target type (neutral ideographs) for liking rating task in the optimal condition. For the valence decision task cases, the moderator target type did not explain much of the heterogeneity, either (𝑄𝑚 =
0.3672, 𝑑𝑓 = 1, 𝑝 = 0.5445). The effect size g = 0.3100 for using emotional faces as targets (k=4, 95% CI= 0.0438, 0.5762, p = 0.0225), g = 0.1657 for using emotional words as targets (k=6, 95% CI= 0.0514, 0.2800, p = 0.0045), g = 0.1876 for using emotional musical pieces (k=1, 95% CI= -0.1612, 0.5363, p = 0.2919). The moderator target type explained 0% of the heterogeneity in the effect sizes (𝜏2 = 0.0299). Still a considerable amount of heterogeneity remained (𝑄𝑒 =
27.1185, 𝑑𝑓 = 15, 𝑝 = 0.0013).
Table 3. Analysis of the secondary moderator target type
Task type Nature of
targets
k Estimate [95% CI] p
Liking rating Neutral ideographs 6 0.4290 [0.2593, 0.5987] <0.0001 Valence decision Emotional faces 4 0. 3100 [0.0438, 0.5762] 0.0225 Emotional words 6 0.1657 [0.0514, 0.2800] 0.0045
Emotional music pieces
1 0.1876 [-0.1612, 0.5363] 0.2919
Suboptimal condition
The effect sizes (k=31) ranged from -0.34 to 2.51, and on aggregate reached medium levels. Testing with the random effects model yielded an overall effect size of g = 0.4354 (p < 0.0001; 95% CI [0.2474, 0.6233]). Most of the effect sizes were bigger than zero as expected (k=26) (see Figure 2).The heterogeneity was estimated to be 𝜏2=0.2523; 95% CI [0.1778, 0.6848]. There was a lot of
heterogeneity in effect sizes in the suboptimal condition (Q=267.0913, df = 30, p < 0.0001). Outlier and influence analysis found one outlier study, g = 2.5061 (studentized residual = -4.9771; Sato & Satoshi, 2006, data of left visual field). Exclusion of this outlier decreased the overall effect size to g = 0.3405 (p < 0.0001; 95% CI = 0.2118, 0.4693). On the other hand, the heterogeneity was decreased (Q = 198.3443, df = 29). All following analyses were conducted under exclusion of this outlier.
Moderator analysis of the suboptimal condition
Three moderator variables (task, negative prime type and method to obtain suboptimality) were included in a mixed effects model. The estimated amount of residual heterogeneity 𝜏2=0.1207;
95% CI [0.0916, 0.4856]. The reduced 𝜏2 suggested 0% of the heterogeneity was explained by the moderator variables tested here (𝑄𝑚= 1.8295, 𝑑𝑓 = 3, 𝑝 = 0.6085). The residual heterogeneity was still high (𝑄𝑒 = 162.8729, 𝑑𝑓 = 26, 𝑝 < 0.0001).
The details of the moderator analysis can be found in Table 4. The test of the moderator variable task explained little heterogeneity (𝑄𝑚 = 0.8420, 𝑑𝑓 = 2, 𝑝 = 0.3588). The average effect size was g=0.4000 for the liking rating task (k= 23, 95% CI [0.2176, 0.5824], p<0.0001) g=0.2223 for the valence decision task (k=7, 95% CI [0.0786, 0.3661], p = 0.0024). The moderator task explained 0% of the heterogeneity in the effect sizes (𝜏2 = 0.1090). Residual heterogeneity
remained large (𝑄𝑒 = 167.9923, 𝑑𝑓 = 28, 𝑝 < 0.0001).
Table 4. Results of moderator analyses of the suboptimal condition
Moderator Level k Estimate [95% CI] p
Task Liking rating 23 0.4000 [0.2176, 0.5824] <0.0001
Valence decision 7 0. 2223 [0.0786, 0.3661] 0.0024 Negative prime sad 11 0. 3549 [0.1984, 0.5115] <0.0001 fearful 8 0. 2898 [0.1150, 0.4646] 0.0012 angry 6 0.1997 [-0.2344, 0.6337] 0.3673 unspecified 5 0.6454 [0.1538, 1.1370] 0.0101
brief presentation 11 0.1873 [-0.0063, 0.3810] 0.0579 CFS 2 0.4596 [-0.0568, 0.9760] 0.0811 Divided attention 1 1.9449 [0.7786, 3.1112] 0.0011
The moderator variable negative prime type explained little heterogeneity (𝑄𝑚 =
0.0424, 𝑑𝑓 = 1, 𝑝 = 0.8369). The moderator task explained 0% of the heterogeneity in the effect sizes (𝜏2 = 0.1118). Residual heterogeneity remained large (𝑄
𝑒 = 186.8016, 𝑑𝑓 = 28, 𝑝 <
0.0001).
The moderator variable masking method did not influence affective priming much )𝑄𝑚 =
0.0206, 𝑑𝑓 = 1, 𝑝 = 0.8859). The moderator task explained 0% of the heterogeneity in the effect sizes (𝜏2 = 0.1101). The test for residual heterogeneity revealed much heterogeneity remaining
(𝑄𝑒 = 190.8482, 𝑑𝑓 = 28, 𝑝 < 0.0001).
Due to the different nature of the liking rating task and the valence decision task, the moderator variable target type was analyzed separately for the two tasks (see Table 5). For the liking rating task cases, the moderator target type explained little of the heterogeneity (𝑄𝑚 =
0.0316, 𝑑𝑓 = 1, 𝑝 = 0.8589). The moderator target type explained 0% of the heterogeneity (𝜏2 = 0.1790). The remaining heterogeneity was almost the same with the heterogeneity before the moderator was considered (𝑄𝑒 = 121.7542, 𝑑𝑓 = 21, 𝑝 < 0.0001). For the valence decision task cases, the moderator target type did not explain much of the heterogeneity, either (𝑄𝑚 =
0.0028, 𝑑𝑓 = 1, 𝑝 = 0.9576). The moderator target type explained 0% of the heterogeneity in the effect sizes (𝜏2 = 0.0299). Again, still a considerable amount of heterogeneity remained (𝑄
𝑒=
37.8425, 𝑑𝑓 = 5, 𝑝 < 0.0001).
Task Target type k Estimate [95% CI] p
Liking rating Neutral faces 7 0.5281 [-0.0437, 1.0998] 0.0703
Nonsense words 1 0.2119 [0.0288, 0.3949] 0.0233 Neutral ideographs 15 0.3815 [0.1918, 0.5713] <0.0001 Valence decision Emotional faces 6 0.2250 [0.0534, 0.3967] 0.0102 Emotional words 1 0.2368 [0.1146, 0.3590] 0.0001 Publication bias
The funnel plot of optimal studies (k=17, Figure 3) appeared to be symmetric. The nonsignificance of Egger’s regression test seemed to confirm this observation, Z=1.6135, p=0.1066, but was more likely to constitute a false negative (Hartgerink et al., 2016). With the trim-and-fill method five cases were filled in at the left side of the funnel plot (see Figure 4), which reduced the average effect size to g = 0.1781; 95% CI [0.0438, 0.3124]. A power check on the basis of the effect size after correction (g = 0.1781) showed that the average power of the 17 effects in the optimal condition was 0.31. This means there should be approximately 5.3 significant effects. The actual number of significant cases was 9, which indicates that there was an excess of three to four significant effects, and so strengthens the conclusion that there was also some publication bias in the optimal affective-priming studies.
Figure 4. Funnel plots for the optimal condition before and after bias correction
The funnel plot of suboptimal studies (k=30, Figure 5) clearly revealed an asymmetric structure. Egger’s regression test supported this observation, Z = 5.4357, p < 0.0001. The trim-and-fill method indicated that three cases should be trim-and-filled in at the left side (see Figure 6). The
correction shifted the effect size to g = 0.285, 95% CI [0.1264, 0.4437]. A power check on the basis of this overall effect size (g = 0.2850) showed that the average power of the 30 effects in the
suboptimal condition was 0.55. With this power, there should be approximately 16.5 significant cases in the total of 30 effects. The actual number of significant cases was 22, which exceeds the theoretical value by 5.5.
Figure 5. Funnel plots for the suboptimal condition before and after bias correction
P-curve method
The P-curve method according to Simonsohn et al. (2014) did not appear to give a reasonable result. The p-curve plots of the two conditions (see Figure 6) are both extremely right-skewed, which, according to the thumb of rule of p-curve method, should be a sign of evidential value. But this does not make much sense, because even cases of very large effect sizes (d > 0.9) without any selection bias did not show such extremely right-skewed plots (Simonsohn et al., 2014). Most likely, the p values were calculated from interaction effects. According to Simonsohn and colleagues (2014), “because when researchers investigate attenuated interactions, p values for simple effects are not uniformly distributed under the null.” “Even if the p-curver is only interested in the simple effect (e.g., because she is conducting a meta-analysis of this effect), that p-value may not be included in p-curve if it was reported as part of an attenuated interaction design. Simple effects from a study examining the attenuation of an effect should not be included in p-curve, as they bias p-curve to conclude evidential value is present even when it is not. (p.543)” Almost all of the p-values were compromised since almost all papers we used do not intentionally focus on the affective priming effect itself, but some other variables. The use of p-curve method did not shed much light on the
Figure 6. P curve plots for both conditions
Dissociation between direct and indirect effects
Twenty-two cases reported effect sizes for the direct effect and were included in the direct-indirect effect graph. The effect sizes for the direct effect ranged from (k=22) -0.3128 to 3.3261. Twenty of these cases were in the suboptimal condition, the overall direct effect size for these twenty cases was g= 0.5433 (k=20, 95% CI = [0.1234, 0.9633], p = 0.0112). The other two cases were in the optimal condition, and the overall direct effect size was g = 2.0836 (k=2, 95% CI = [1.7203, 2.4470], p < 0.0001).
The indirect effects were plotted against the direct effects in Figure 7. At least a simple dissociation seemed present in these affective priming studies. When the direct effect was at chance level, the indirect measure still had a robust above zero positive value, indicating the involvement of nonconscious processing in affective priming. The plot might also reveal a double dissociation, as it can be seen that as the direct measure effect increased, the indirect measure effect decreased. This again confirmed the existence of nonconscious processing independent from conscious processing, and indicated that, as the direct effect increased, the nonconscious information sources that fueled the indirect effect were inhibited, because all the dots at the right side of the plot lied in
Figure 7. Mapping of indirect effects against the direct effects
Publication bias of the direct effect sizes
The funnel plot of the suboptimal studies (k=20, Figure 8) appeared to be obviously asymmetric, but Egger’s regression test suggested there was no significant bias, Z=0.0290, p=0.9769. With the trim-and-fill method, however, three cases were filled in at the right side of the funnel plot (see Figure 8), which increased the average effect size to g = 0.5479; 95% CI [0.2155, 0.8804]. It should be noted that in contrast to regular publication bias there was a bias here against publishing
significant, instead of nonsignificant results. This finding clearly supported the conclusion of Vadillo et al. (2015) that the indirect-without-direct effect was largely based on false negatives.
Figure 8. Funnel plots for the effect sizes of the direct effect before and after bias correction
Discussion
Aggregated over all studies, the meta-analysis yielded clear affective priming, but more insight into the underlying processes can only be gained by further analyzing publication bias, moderators, and other constraining factors. Both funnel plots of the two conditions suggested there were too many cases at the right side of the funnel plots. The application of the trim-and-fill method filled in cases at the left side of the funnel plots for both conditions, and even reduced the optimal effect size to a larger extent than the suboptimal effect size. This suggested that publication bias occurred in both conditions, not just for suboptimal conditions, as was expected. Therefore, the parallel shifts of the effect sizes to the left did not change much of the relative difference between these two conditions. The meta-analysis found a robust small-to-medium overall affective priming effect in both optimal and suboptimal conditions. It can be concluded that affective priming effect is present no matter whether the experiment is set in an optimal or suboptimal condition. This contradicts the seminal finding of Murphy and Zajonc (1993), in which they found that affective priming is absent in the optimal condition.
The affective-priming effect was slightly larger in the suboptimal condition than in the optimal condition. This is consistent with studies supporting a suboptimal advantage in affective
priming. Although the affective priming did not disappear in the optimal condition, it was
somewhat reduced. This contradicts the assumption that nonconscious processing is a diluted form of conscious processing. The larger suboptimal than optimal affective-priming effect thus suggests a qualitative difference between the two types of processing. This qualitative difference is also revealed by the direct vs. indirect effect graph, in which not only an indirect-without-direct effect but even a double dissociation could seemingly be observed (cf, Schmidt & Vorberg, 2006).
It should be noted, however, that the typical double dissociation in the plot can be a result of a few large effects with high variance at the zero-direct effect axis. Similar high variance effects did not occur at the low end of the indirect-effect scale, and are thus probably indicative of publication bias. Such a publication bias was indeed observed and corrected for both in suboptimal and optimal conditions, but could not be corrected in the direct-indirect graph. After removal of these high variance effects, the slope of indirect effect changed along with the increase of the direct effect seems much flatter and the double dissociation became less apparent typical.
Publication bias, in reverse direction against publishing significant effect, may also have affected the direct effect. The funnel plots and the trim-and-fill method indicated that the direct effect was actually underestimated in the suboptimal condition. This underestimation might have been a result of false negatives due to low power, as well as of selective reporting. To correct the direct effect by shifting many of the suboptimal cases to the right in the D-I plot, the evidence for a double dissociation became more diffuse and actually diminished. This line of reasoning does not exclude the possibility of a suboptimal advantage, but even if it does exist, it should remain small.
The occurrence of the stronger-suboptimal-than-optimal affective priming pattern may be mediated, and even confounded, by other factors than the manipulation of conscious processing. The moderator variables chosen for the meta-analysis could explain some of the heterogeneity. In both optimal and suboptimal conditions, studies using sad faces as negative primes yielded
comparable medium-sized effects, but those using fearful and angry faces only yielded small effect sizes. Sad faces thus appeared best suitable for obtaining affective priming. This was not what we
expected if the affective meaning of anger is more unambiguous than that of the fearful and sad faces. A possibility is that negative expressions differ in the processing strategies they evoke in an implicit manner. Sad faces may, for instance, elicit more nonanalytic processing tendencies, whereas fearful and particularly angry faces may raise analytic tendencies. Gable and Harmon-Jones (2010) have indeed found that sad moods, which is low in motivational intensity, direct attention away from the details to a more global (i.e., nonanalytic) processing strategy, whereas disgust, which is high in motivational intensity, narrowed attentional processing to the details (i.e., resulted in analytic processing). If the adoption of a non-analytic processing strategy enhances priming, stronger priming should be expected for sad faces than for fearful and angry faces,
irrespective of presentation condition. This explanation thus suggests that the (implicit) adoption of processing strategy matters more in affective priming than the optimal-suboptimal manipulation.
For another moderator variable, target type, the result was also somewhat unexpected. The sample size for the valence-decision task was too small to indicate anything important, but for the studies using the liking rating task, neutral faces did a better job in eliciting priming than neutral ideographs. This again, might be the result of different processing strategies evoked by faces and ideographs. If the following stimuli shared some similar features with the preceding stimuli (in this case, both the primes and targets were faces), a global and non-analytic processing strategy may be more likely to be selected by the participants. Also the task to rate neutral stimuli in the liking rating task seems to rely more on non-analytic processing than determining whether a stimulus is
positively or negatively valenced in the valence-decision task. Determining the latter affective meanings requires discriminating specific stimulus characteristics and therefore presumably involves more analytic processing than globally evaluating neutral stimuli.
For the moderator variable method to obtain suboptimality, it seems backward masking did a better job than brief presentation. This could be expected from a consciousness view, because the masking may further reduce conscious processing than brief presentation only. However, backward masking did not prove superior to CFS method, which might result from the limited number of
cases in the CFS condition. The results with this moderator variable can also be explained from the perspective of processing strategy. The manipulation of dividing attention and the loading of working memory may be seen to hamper analytic processing, because they reduce controlled processing which is needed for such an analysis. The methods of masking blur the prime details, and for this reason make analytic processing implausible. The brief-presentation method, finally, would seem to leave the most room for analytic processing, and would thus be associated with the smallest priming effects.
The moderator task explained some of the heterogeneity for both conditions, especially for the optimal condition. In the liking rating task, the effect sizes reached a medium level, but for the valence decision task, the effect sizes remained small. This might cast some doubt on the
conclusion that suboptimal affective priming would be stronger than optimal affective priming: in the suboptimal condition, 77% of the effect sizes came from studies using the liking rating task, whereas in the optimal condition 65% of the effect sizes came from valence decision studies. The effect size difference between these two conditions can purely be a result of different task
frequencies, which could of course not be controlled in the meta-analysis. In fact, if we compare two conditions with the confounding of task eliminated (see Table 6), the effect sizes in the optimal condition were similar to those in the suboptimal condition when separated for liking rating and valence decision. The overall effect size in the suboptimal condition may thus have seemed bigger only because it had more liking rating studies.
Table 6. Effect sizes g of affective priming effect in the contingency table of task and optimality Task
Optimality
Liking rating task (N) Valence decision task (N)
Optimal 0.4290 (7) 0.2263 (11)
Compared to the suboptimal condition, the heterogeneity was smaller in the optimal condition. At first, it was believed that this is because in the optimal condition masking method would not be a moderator adding heterogeneity into the effect sizes. However, the moderator analysis of the suboptimal condition did not reveal much heterogeneity arising from the masking method. There must be some other moderator variable, such as processing strategy, operating. A possibility is, the participants are more likely to use variable processing strategies in a relatively vague situation, as in this case in the suboptimal condition. In the optimal condition, in which specific prime features can be consciously distinguished, participants are more likely to adopt an analytic approach, thereby reducing variability. The enhanced heterogeneity in suboptimal
conditions, thus further supports that a hidden crucial moderator variable confounds the suboptimal-optimal dissociation, but that this variable could not be captured in the present meta-analysis, due to insufficient studies being available explicitly manipulating processing strategy.
Apart from processing strategy, there may also be individual differences mediating the strength of suboptimal-optimal affective priming. Lapate and collaborators (2016) recently suggested that the activity of amygdala and the fiber connection thickness between amygdala and ventromedial PFC (i.e., the uncinate fasciculus) could moderate the affective priming effects in suboptimal and optimal conditions, respectively. Comparing priming by fearful faces and flowers in a liking rating task, they found no significant affective priming, either in unaware or aware
conditions. Over participants the priming correlated positively with right amygdala activation (r = 0.40) in the former condition, suggesting that only participants showing an amygdala sensitivity for fearful faces may reveal suboptimal affective priming. The inhibition of affective priming in the aware condition depended strongly on the inverse coupling between PFC and amygdala. Only in participants with such a strong coupling should the stronger-suboptimal-than-optimal priming occur. Analytic processing is likely to be associated with PFC activity, so these results again argue for processing strategy as the mediating factor.
The field of affective priming by faces does not seem to have advanced sufficiently to draw stronger conclusions about when the effect can be found and when not. The discussion above however points to a potentially important role of processing strategy in boosting and diminishing the affective priming effect, similar to what Whittlesea and Price (2001) have found for the mere-exposure effect. Future studies should certainly look into this factor to check whether this mediation assumption holds true. In sum, the effect sizes for affective priming obtained in this meta-analysis remained rather small and did seem to differ for both presentation conditions. Small effects are associated with a large deal of nonreplications, which may explain the difficulty in finding this type of affective priming. Processing strategy, which in many studies may be confounded with
manipulation of conscious processing, seems the most likely candidate for increasing affective priming, irrespective of the level of conscious processing.
Acknowledgments
We thank Wataru Sato, Satoshi Aoki, Mikko Lähteenmäki, Lauri Nummenmaa, Jukka Hyönä and Mika Koivisto for kindly offering the supplementary data of their studies to us. We also thank Jelte Wicherts for providing the Excel table to estimate the correlation coefficient.
References1
Adams, W.J., Gray, K.L., Garner, M., & Graf, E.W. (2010). High-level face adaptation without awareness. Psychological Science, 21, 205-210.
*Aguado, L., Dieguez-Risco, T., Méndez-Bértolo, C., Pozo, M. a, & Hinojosa, J. a. (2013). Priming effects on the N400 in the affective priming paradigm with facial expressions of emotion. Cognitive, Affective & Behavioral Neuroscience, 13, 284–296.
Alexopoulos, T., Fiedler, K., & Freytag, P. (2012). The impact of open and closed mind sets on evaluative priming. Cognition & Emotion, 26, 978-994.
Alexopoulos, T., Lemonnier, A., & Fiedler, K. (2017). Higher order influences on evaluative priming: Processing styles moderate congruity effects. Cognition and Emotion, 31, 57-68. *Almeida, J., Pajtas, P. E., Mahon, B. Z., Nakayama, K., & Caramazza, A. (2013). Affect of the
unconscious: visually suppressed angry faces modulate our decisions. Cognitive, Affective & Behavioral Neuroscience, 13, 94–101.
Anderson, E., Siegel, E., White, D., & Barrett, L. F. (2012). Out of Sight but Not Out of Mind: Unseen Affective Faces Influence Evaluations and Social Impressions. Emotion, 12, 1210– 1221.
Barrett, L.F., Mesquita, B., & Gendron, M. (2011). Context in emotion perception. Current Directions in Psychological Science, 20, 286-290.
Blaison, C., Imhoff, R., Huhnel, I., Hess, U., & Banse, R. (2012). The affect misattribution procedure hot or not? Emotion, 2, 403-412.
Carver, C.S., & Harmon-Jones, E. (2009). Anger is an approach-related affect: Evidence and implications. Psychological Bulletin, 135, 183-204.
Cesario, J. (2014). Priming, replication, and the hardest science. Perspectives on Psychological Science, 9, 40-48.
*Chiesa, P. A., Liuzza, M. T., Acciarino, A., & Aglioti, S. M. (2015). Subliminal perception of others’ physical pain and pleasure. Experimental brain research, 233, 2373-2382. *Comesaña, M., Soares, A. P., Perea, M., Piñeiro, A. P., Fraga, I., & Pinheiro, A. (2013). ERP
correlates of masked affective priming with emoticons. Computers in Human Behavior, 29, 588–595.
Cooper, H., Hedges, L. V., & Valentine, J. C. (Eds.). (2009). The handbook of research synthesis
and meta-analysis. New York, NY: Russell Sage Foundation.
Chan, R. C. K., Li, H., Cheung, E. F. C., & Gong, Q.-Y. (2010). Impaired facial emotion perception in schizophrenia: a meta-analysis. Psychiatry Research, 178, 381–390.
Dannlowski, U., Ohrmann, P., Bauer, J., Kugel, H., Arolt, V., Heindel, W., & Suslow, T. (2007). Amygdala reactivity predicts automatic negative evaluations for facial emotions. Psychiatry Research - Neuroimaging, 154, 13–20.
Desender, K., & Van den Bussche, E. (2012). The magnitude of priming effects is not independent of prime awareness. Reply to Francken, van Gaal, & de Lange (2011). Consciousness and Cognition, 21, 1571–1572.
*Donges, U. S., Kersting, A., & Suslow, T. (2012). Women’s greater ability to perceive happy facial emotion automatically: Gender differences in affective priming. PLoS ONE, 7, 1–5. Dunlap, W. P., Jose, J. M., Vaslow, J. B., & Burke, M. J. (1996). Meta-analysis of experiments with
matched groups or repeated measures designs. Psychological Methods, 1, 170–177. Egger, M., Smith, G. D., Schneider, M., & Minder, C. (1997). Bias in meta-analysis detected by a
simple, graphical test. Bmj, 315, 629-634.
Gable, P., & Harmon-Jones, E. (2010). The blues broaden, but the nasty narrows attentional consequences of negative affects low and high in motivational intensity. Psychological Science, 21,211-215.
Goodkin, O., & Phillips, M. A. (1980). Subconscious Taken Captive: A Social, Ethical, and Legal Analysis of Subliminal Communication Technology, The Southern California Law
Review, 54, 1077.
Greenwald, A. G., Spangenberg, E. R., Pratkanis, A. R., & Eskenazi, J. (1991). Double-blind tests of subliminal self-help audiotapes. Psychological Science, 2, 119–122.
Guinote, A., Willis, G. B., & Martellotta, C. (2010). Social power increases implicit prejudice. Journal of Experimental Social Psychology, 46, 299–307.
Hartgerink, C. H. J., van Assen, M. A. L. M., & Wicherts, J. M. (2016). Too good to be false: Non-Significant results revisited. Manuscript to be published.
Hooker, C. I., Tully, L. M., Verosky, S. C., Fisher, M., Holland, C., & Vinogradov, S. (2011). Can I trust you? Negative affective priming influences social judgments in schizophrenia. Journal of Abnormal Psychology, 120, 98–107.
Iacono, V., Ellenbogen, M. A., Wilson, A. L., Desormeau, P., & Nijjar, R. (2015). Inhibition of personally-relevant angry faces moderates the effect of empathy on interpersonal functioning. PLoS ONE, 10. http://doi.org/10.1371/journal.pone.0112990
Ioannidis, J. P. A. (2005). Why most published research findings are false. PLoS Medicine, 2, 0696–0701. http://doi.org/10.1371/journal.pmed.0020124
*Inuggi, A., Sassi, F., Castillo, A., Campoy, G., Leocani, L., Garc??a Santos, J. M., & Fuentes, L. J. (2014). Cortical response of the ventral attention network to unattended angry facial
expressions: An EEG source analysis study. Frontiers in Psychology, 5, 1–13. http://doi.org/10.3389/fpsyg.2014.01498
*Jiang, J., Bailey, K., Chen, A., Cui, Q., & Zhang, Q. (2013). Unconsciously Triggered Emotional Conflict by Emotional Facial Expressions. PLoS ONE, 8, 1–7.
http://doi.org/10.1371/journal.pone.0055907
*Jostmann, N. B., Koole, S. L., Van Der Wulp, N. Y., & Fockenberg, D. A. (2005). Subliminal affect regulation : The moderating role of action vs. State orientation. European Psychologist, 10, 209–217.
*Kamio, Y., Wolf, J., & Fein, D. (2006). Automatic processing of emotional faces in
high-functioning pervasive developmental disorders: An affective priming study. Journal of Autism and Developmental Disorders, 36, 155–167.
*Kamiyama, K. S., Abla, D., Iwanaga, K., & Okanoya, K. (2013). Interaction between musical emotion and facial expression as measured by event-related potentials. Neuropsychologia, 51(3), 500–505.
