Approach, avoidance, and affect

Approach, avoidance, and affect: A

meta-analysis

Research Report

Sören Mohr

Abstract

The automaticity of the link between affective information processing and approach and avoidance action tendencies was examined in a meta-analysis on 29 published studies. There is considerable debate whether positively and negatively valenced stimuli prime approach and avoidance movements directly, or whether this link is mediated by interpretations of these movements and intentions to react to these stimuli. Instruction, stimulus type, and type of task served as moderator variables. A significant small to

medium-sized overall effect was found for both positive and negative affect. Stimulus type and task yielded no significant differences in effect size. Explicit instructions referring to the self or the stimulus object yielded larger effect sizes than implicit instructions. The link between affect and approach-avoidance seems to be dependent on intentions to deal with the affectively laden stimuli.

Introduction

Emotions concern personally relevant events that in many cases demand urgent responses. In our daily lives, we are often faced with situations that call for quick and appropriate action. Grasping opportunities to obtain a job or avoiding unsafe places at night in big cities are essential behaviors driven by strong emotions. Many emotion theories postulate a fundamental link between emotions and primitive automatic action tendencies, such as, for instance, approach and avoidance (e.g., Frijda, 1986). Emotions are sometimes assumed to be organized into two different motivational systems that prepare the organism to respond appropriately to emotionally significant stimuli in the environment (Lang, Bradley, & Cuthbert, 1990). Appetitive motivational circuits direct the organism to approach positively valenced stimuli, whereas a defensive motivational system serves to trigger avoidance behavior away from negatively appraised stimuli. In line with this theorizing, a seminal study by Solarz (1960) showed that perceiving stimuli with a positive valence fosters approach behavior, whereas negative stimuli facilitate avoidance behavior. Participants saw pleasant and unpleasant words on cards that were fixed to a movable stage. They were faster to pull cards with pleasant words toward themselves and to push cards with unpleasant words away from themselves. This compatibility effect has been replicated many times with similar paradigms and a large range of affective stimuli (e.g., Chen & Bargh, 1999; De Houwer, Crombez, Baeyens & Hermans, 2001; Rinck & Becker, 2007), supporting a link between affective information processing and approach and avoidance behavior.

Specific emotional (i.e., evolutionary prepared) stimuli may be picked up very quickly and receive processing priority (Öhman, 1986), influencing subsequent behavior even if they are not perceived fully consciously (Rotteveel, Groot, Geutskens, & Phaf, 2001). It is plausible to assume that much affective information processing is automatic. Emotional stimuli can be evaluated automatically on a positive-negative affect dimension

(i.e., affective primacy; Zajonc, 1980). Chen and Bargh (1999) further claimed that

automatic evaluation of stimuli in turn automatically predisposes approach and avoidance reactions to them. In their experiment, participants were instructed to pull a lever towards themselves (i.e., approach) or to push it away from themselves (i.e., avoidance) regardless of the valence of the stimuli. Compatibility effects were found even when participants were not explicitly instructed to evaluate the meaning of the presented words. The authors interpreted their finding as demonstrating a direct, automatic link between motivational states of approach and avoidance and specific motor actions (i.e., arm flexion and extension). Other studies, using a computer joystick as the dependent measure seem to support this claim. Participants are mostly instructed to evaluate an irrelevant feature of valenced stimuli (e.g., the background color). However, many of these studies also provided visual feedback (cf. Seibt, Neumann, Nussinson, & Strack, 2008). Pulling the joystick increased, whereas pushing the joystick decreased the size of the stimuli. It can be argued that the zooming feature explicitly reinforces the interpretations of the

respective arm movements, which puts the automatic link between affect and arm movements in question.

Also other research suggests that the link between affect and approach-avoidance may not be as automatic as Chen and Bargh claim. Rotteveel and Phaf (2004), for

instance, used a vertical stand with three buttons to measure approach and avoidance behavior. Pressing the upper button and the lower button corresponds to arm flexion and extension, respectively. When participants were instructed to evaluate an irrelevant feature (i.e., gender of emotional faces), no compatibility effects were found. The instructions may not have induced the participants to affectively label these flexion and extension

movements. It is still unclear, whether action tendencies for arm flexion and extension as a result of affective information processing depend on the intention to evaluate the affective meaning of stimuli.

The notion that flexor and extensor movements are associated with approach- and avoidance-related motivations is also supported by a study from Cacioppo, Priester, and Berntson (1993). Flexor and extensor movements differentially modulated participants' preferences for neutral ideographs. On theoretical grounds, the authors argue that arm movements become conditionally associated with approach and avoidance motivations. Flexion most often becomes associated with the retrieval or ingestion of something desired, whereas extension is mostly coupled with pushing away something aversive (Maxwell & Davidson, 2007).

On the other hand, it is easy to conceive of situations where the function of flexor and extensor movements is reversed. Depending on the particular context, the same physical movement has different effects. Seeing a spider should elicit the reflex to withdraw the hand, especially for individuals suffering from a spider phobia. In addition, approach movement can go along with reaching out one's hand to pet a beloved animal. Seibt, Neumann, Nussinson, and Struck (2008) supported this theorizing by inducing an object-related frame of reference. By instructing participants to move the joystick towards or away from the word on the screen, they found opposite compatibility effects. Numerous studies have found similar effects that seem to contradict interpretations in terms of a hard-wired relationship between approach-avoidance motivations and particular arm

movements (e.g., Eder & Rothermund, 2008; Lavender & Hommel, 2007; Markman & Brendl, 2005). The phrasing of instructions seems to have a large influence on how particular movements of the arm are interpreted by participants.

Evaluative-response-coding accounts (Eder & Rothermund, 2008) even go as far as to claim that valence also has no special status among other stimulus features, such as size, color, and location. Approach and avoidance behaviors are seen to follow general principles of action control, instead of being regulated by distinct motivational mechanisms (Eder & Rothermund, 2008; Lavender & Hommel, 2007). According to this view,

compatibility effects are due to a match between evaluative codings of approach and avoidance movements and the valence of the stimuli. There is empirical evidence that situational demands influence the meaning of arm movements, casting doubt on the existence of fixed effects of biceps and triceps activations. However, the commonly used joysticks and levers not only involve biceps and triceps muscles but also pulse and even shoulder muscles, which may be less consistently related to approach-avoidance than biceps and triceps. In contrast, the vertical button stand could be considered a purer or more implicit measure of arm flexion and extension, because instructions are usually not contaminated with references to approach and avoidance behaviors (i.e. toward and away) and the flexion and extension do not involve muscle movements of the hand. Furthermore, the vertical stand does not move away or towards the self or an object, holding the

distance constant between the self and an object.

So, in addition to how the instructions are phrased, comparing different approach-avoidance tasks can shed light on the main question of this meta-analysis: is there an automatic link between affective information processing and action-tendencies of approach and avoidance?

Krieglmeyer and Deutsch (2010) provided a direct comparison of often-used measures of approach-avoidance behavior, but the field lacks a quantitative review of the available data. Divergent research results have also given rise to the assumption that several factors moderate the effect of affective information processing on approach-avoidance behaviors. The goal of the current meta-analysis is to shed light on the influence of potential moderators and to provide an estimation of the overall effect size. The meta-analysis is focused on studies using arm movements and abstract approach-avoidance behavior as the dependent measure. The following section provides a description of the moderators included.

Tasks

Research suggests that approach-avoidance tasks differ in their sensitivity to valence (Krieglmeyer & Deutsch, 2010). One type of task only measures

approach-avoidance behavior on an abstract level, not involving any arm movements (De Houwer et al., 2001). Participants control a manikin on the computer screen that appears randomly above or below a stimulus. By means of key presses they either move the manikin towards or away from the stimulus. This task will here be referred to as the abstract manikin task.

In the joystick task, approach-avoidance behavior is operationalized as horizontally pulling or pushing a vertically positioned control stick. This may involve flexion and

extension of the arm, but also pulse and shoulder movements. The same applies to a lever when used as approach-avoidance task (Chen & Bargh, 1999). These measures are therefore treated as the same task.

The feedback-joystick task, however, is to be considered a separate task. It was shown that the visual feedback is resistant to cognitive reinterpretations (Rinck & Becker, 2007). When a stimulus-reference point was induced by rephrasing the instructions (i.e., pull the joystick away from the picture, push the joystick towards the picture), the

compatibility effect was not reversed. Therefore, the crucial aspect in the feedback-joystick task seems to be the visual impression that the stimuli come closer or disappear.

Another distinct measure is the vertical three-button stand (Rotteveel & Phaf, 2004). Pressing the upper button and the lower button corresponds to arm flexion and extension, respectively. Participants first press the middle button and are simply told to press the upper or lower button when the stimulus appears. Here movements are made in the

vertical direction by either flexing the arm with the biceps muscle or extending the arm with the triceps muscle. No other muscles are engaged in this task. Instructions are usually not contaminated with explicit references to approach and avoidance behaviors (i.e., toward and away).

For all tasks, compatibility effects are defined as positive stimuli facilitating approach movements and negative stimuli facilitating avoidance behavior. Reaction times are

measured from the onset of the stimulus. Only the three-button stand provides two dependent measures: the release time of the middle button and the movement time to reach one of two response buttons. The current meta-analysis only considered the release time which should reflect the influence of affect on preperation times (Rotteveel & Phaf, 2004).

By comparing different tasks, it will be investigated how well they represent automatic approach-avoidance behaviors and to what extent they are influenced by cognitive labeling of these actions.

Instructions

A central question in the literature is whether affective information processing automatically triggers approach-avoidance responses, independent of the intention to evaluate the stimuli. Three different instructions were compared to investigate this question. With explicit instructions participants are instructed to evaluate the stimuli with the

approach-avoidance task on a positive-negative dimension. They are, for instance, told to pull the joystick towards them or push it away from them when they judge the stimulus as positive or negative, respectively. Implicit instructions do not require participants to attend to the valence of the stimuli. Instead, they are instructed to react to a task-irrelevant feature (e.g., the background color or the gender of a face). The most extreme form of implicitness can be found in research where also the affective valence of the stimuli is implicit (i.e., not consciously recognized by the participant). There are, however, only very few examples of such studies (e.g., Jones, Young, & Claypool, 2011; Phaf & Rotteveel, 2009). Explicit instructions usually yield larger effect sizes than implicit instructions (e.g., Krieglmeyer & Deutsch, 2010).

The third type of instructions will be termed incongruent. Here, the meaning of flexion and extension is reversed, usually by changing the reference point of arm

movements. Flexion of the arm now corresponds to avoidance, whereas extension reflects an approach movement. Incongruent conditions have only been tested in combination with explicit instructions, which means that participants were instructed to respond to the

valence of stimuli.

By comparing explicit and incongruent conditions, the question can be addressed whether action-tendencies to approach and avoid are context-independent movements consisting of specific motor patterns (arm flexion and extension). Further, effects for implicit instructions would argue for an automatic link between affective information processing and approach-avoidance behavior that does not depend on the intention to evaluate the affective meaning of stimuli.

Stimuli

Specific stimuli might be more suitable than others to automatically induce affect and thus to measure approach-avoidance behavior. In many studies, the same participants were tested with both positive and negative stimuli. This induces interdependence

between reaction times. In order to extract multiple effect sizes from the same study, positive and negative stimuli were analyzed separately. In addition, some studies only reported reaction times pooled across compatible (i.e., approach positive and avoid

negative stimuli) and incompatible trials (i.e., avoid positive and approach negative stimuli), respectively. Therefore, an additional analysis was conducted for pooled reaction times. Within each analysis, four types of stimuli were investigated. The first type of studies used words with an emotional content as stimuli. These are words that have been selected on the basis of their strong affective valenced and thus can be explicitly evaluated on a positive-negative dimension.

Furthermore, we considered pictures depicting emotional scenes, mostly selected from the International Affective Picture System (IAPS) (Lang, Bradley, & Cuthbert, 1996).

Emotional facial expressions are presumably evolutionary prepared stimuli and might therefore be processed more automatically (Öhman, 1986) than for instance words, which may differ widely across languages. To ensure comparability across studies, only happy and angry facial expressions were included.

Tests of approach-avoidance behavior are often used to assess action-tendencies towards personally relevant stimuli. These stimuli are predominantly spider pictures that are tested with participants suffering from a spider phobia.

Method

Search Procedure

A literature search for relevant studies was conducted across four databases (ISI Web of Science, PsycINFO, PubMed, and Google Scholar) using the search string

“approach-avoidance behavior or approach-avoidance task or compatibility”. The search in PsycINFO resulted in 325 references. In addition, cited reference searches were

conducted in ISI Web of Science to search for studies that referred to the following studies: Chen & Bargh, 1999 (283 references); Rinck & Becker, 2007 (41 references); Rotteveel & Phaf, 2004 (49 references). Additional studies were identified by manual search.

Inclusion and Exclusion Criteria

Studies were included according to the following criteria: (1) Studies investigated healthy participants. (2) No studies involving longer-term emotional moods were included. Data from control conditions and behavioral assessments prior to a to-be-excluded

manipulation were, however, included (Roelofs, Elzinga, & Rotteveel, 2005). (3) Studies employed either the joystick/lever, feedback-joystick, abstract tasks or the button stand as

the dependent measure. This resulted in the exclusion of studies that had whole-body movements as the dependent measure (e.g., Stins, Roelofs, Villan, Kooijman, Hagenaars & Beek, 2011), or that investigated the reverse effect of arm movements on evaluation of stimuli (e.g., Cacioppo et al., 1993). (4) Studies that reported relevant means and standard deviations (or standard errors). Studies that did not report these statistics were excluded, unless authors could be contacted to provide them.

All studies were published between 1999 and 2012. Altogether, the meta-analysis included 29 studies, from which 81 effect sizes were obtained (Combined N = 1538). A detailed overview of the studies on the effect of affective information processing on approach-avoidance behaviors is provided in Appendix A through Appendix C.

Effect Size Calculation

In the majority of studies, effect sizes could be computed on the basis of means and standard deviations. Effect sizes were computed manually in terms of Cohen's d (see Formula 1). pooled comp inc

S

M

M

d





It refers to the standardized mean difference between experimental conditions, an incompatible condition (INC) and a compatible condition (COMP), divided by the pooled standard deviation (Hedges & Olkin, 1985) (see Formula 2). A compatible condition refers to a situation where subjects had to approach positive stimuli or avoid negative stimuli. With explicit instructions and a task involving arm movements, approach corresponds to arm flexion and an avoidance corresponds to arm extension. In the incompatible condition, subjects approached negative stimuli or avoided positive stimuli. With incongruent

instructions, the coupling between valence and the flexor and extensor movements is usually reversed. In the study by Eder and Rothermund (2008; Experiment 3), left and right

movements were labeled positively and negatively, respectively. These instructions were also coded as incongruent instructions in the current meta-analysis, in the sense that this response-label assignment is incongruent with what we refer to as explicit instructions.









2

²

1

²

1













n

n

S

n

S

n

S

Cohen's d has a slight bias, tending to overestimate the absolute population value of the effect size in small samples. Cohen's d can be converted to Hedges' g, using the

correction factor J (see Formula 3).

1

4

3

1







J

df refers to the degrees of freedom, which is n – 1, where n is the number of pairs.

Hedges' g is then computed by multiplying Cohen's d by the correction factor J,

J

d

g





This unbiased estimator, Hedges' g (Hedges & Olkin, 1985), was used for subsequent analyses.

Because the majority of studies had repeated measures designs (k = 26), the variance of g was given by Formula 4.



1



²

2

2

²

1

J

r

n

d

n

v













 



r is the correlation between pairs of observations. Because this correlation strongly

influences the variance of g, it should not be ignored (see Dunlap, Cortina, Vaslow & Burke, 1996).

For those studies using an independent groups (i.e., between-subjects) design, the correction factor J was calculated according to Formula 3. df is here equal to n1 + n2 – 2,

and n1 and n2 are the sample sizes in the two groups. The variance of g for independent





²

2

²

J

n

n

d

n

n

n

n

v









All studies in the current meta-analysis that used independent groups had equal sample sizes per group.

Missing Data

Whenever necessary, authors were contacted to gather means and standard deviations to compute effect sizes. Unfortunately, it is not common practice to report r, the correlation between pairs of observations in repeated measures designs. Therefore, r was estimated from paired t-tests according to Formula 6.

²

2

²

²

2

t

n

g

t

r





It was calculated from repeated measures ANOVAs according to Formula 7.

F

n

g

F

r

2

²

2





For some studies (Phaf & Rotteveel, 2009; Seidel, Habel, Finkelmeyer, Schneider, Gur, & Derntl, 2010; Seidel, Habel, Kirschner, Gur, & Derntl, 2010), r could be computed from the raw data and compared to estimations derived from test-statistics. The estima-tions turned out to be fairly accurate, validating the use of the formulas mentioned above. For the remaining studies, that did not report the relevant test-statistics, the average of all available correlations, weighted by individual sample sizes, was imputed as the correlation for the individual study. If means and standard deviations are not provided, and if the cor-relation between measures is not reported nor can be estimated appropriately, then it is recommended to exclude the study (Dunlap, Cortina, Vaslow, & Burke, 1996).

Data Analysis

All analyses were computed in the statistical software package R (version 2.14.1) (R Development Core team, 2010) by using the metafor package (Viechtbauer, 2010). Due to initial heterogeneity, all analyses were computed within the random effects model. The proportion of systematic unexplained variance (tau²) was estimated using maximum-likelihood estimation, which is approximately unbiased and quite efficient (Viechtbauer, 2010). Cochran’s Q-test (Hedges & Olkin, 1985) was used to test homogeneity of the effect sizes.

Outlying cases were identified based on standardized residuals. Influence analysis (i.e., the exclusion of single studies) was performed to identify influential studies based on the Cook's distance and residual heterogeneity.

Moderating variables were defined a priori. Hypothesized categorical moderators were (1) task (vertical button stand, joystick/lever, feedback-joystick, abstract manikin task), (2) instruction (explicit (task-relevant), incongruent (task-relevant), implicit (task-irrelevant)), (3) stimulus type (emotional facial expressions, emotional words, emotional pictures,

personally relevant stimuli), (4) design (repeated measures design, independent groups design), (5) valence (explicitly valenced stimuli, implicitly valenced stimuli). The statistic

Qm was used as an omnibus test for differences between levels.Qe was used to test for

residual heterogeneity.

Publication Bias

An important issue in meta-analyes is the possibility of publication bias. Studies with statistically significant effects and positive treatment outcomes are more likely to be

published than null results (Begg, 1994). If a publication bias is present, the studies included in the meta-analysis are not representative of all valid studies undertaken in the field, leading to an over-estimation of the effect. If studies with non-significant results

remain unpublished, this may be reflected in an asymmetric funnel plot. In the present meta-analysis, the possibility of a publication bias was tested by conducting a regression test for funnel plot asymmetry (Egger, Smith, Schneider, & Minder, 1997). To correct for a possible publication bias, the trim-and-fill method was applied (Duval & Tweedie, 2000). This method estimates the number of missing studies and provides an adjustment of the overall effect size.

Results

Positive Affect

Figure 1 shows the funnel plot for the analysis of positive affect (k = 27). As can be seen, the majority of the effect sizes were in the expected direction (k = 25). Twelve of these positive effect sizes were significant. Two studies showed an effect in the opposite direction, one of which was significant. The effect sizes ranged from g = -0.08 to g = 1.29. Influence analysis identified three outliers, g = 1.29 (standardized residual = 0.9832)

(Markman & Brendl, 2005; A), g = 1.06 (standardized residual = 0.7545) (Markman & Brendl, 2005; B), g = 0.8 (standardized residual = 0.5012) (Phaf & Rotteveel, 2009; Experiment 2A).

The random effects model yielded a significant average effect size (g = 0.3067; p < 0.0001; 95% CI = 0.2000, 0.4135). The estimated amount of heterogeneity was equal to

tau² = 0.0572; 95% CI = 0.0293, 0.1478. There was a clear indication of heterogeneity in

effect sizes (Q = 183.2421, df = 26, p < 0.0001). The exclusion of the three outliers resulted in a reduced average effect size (g = 0.2164; p < 0.0001; 95% CI = 0.1413,

0.2915). The unexplained variance component was diminished (tau² = 0.0172), but the test of heterogeneity was still highly significant (Q = 93.0283, df = 23, p < 0.0001). All

moderator analyses were conducted under exclusion of the three outliers (except for the analysis of the moderator valence, because one outlier (Phaf & Rotteveel, 2009; A) was

the only study using stimuli with implicit valence.

Figure 1. Funnel plot of effect sizes for positive affect, representing a scatterplot of

treatment effect on the x-axis against the standard error on the y-axis (see Appendix A).

Moderator Analyses of Positive Affect

Four moderator variables (task, instruction, stimulus type, design) were included in a mixed effects model. The estimated amount of residual heterogeneity was equal to tau² = 0.0000; 95% CI = 0.0000, 0.0044, suggesting that at least 74% of the variance in effect sizes could be accounted for by including the moderators (Qm = 83.4099, df = 7, p <

0.0001). The test for residual heterogeneity was not significant (Qe = 9.6185, df = 16, p =

0.8858). It is worthwhile to consider each moderator separately.

The results of the moderator analyses are provided in Table 1. The test of the moderator task was significant (Qm = 6.5402, df = 2, p = 0.038). The average effect size

was significantly different from zero for the vertical stand (g = 0.2721; p = 0.0008; 95% CI = 0.1127, 0.4316), as well as for the joystick/lever (g = 0.2509; p < 0.0001; 95% CI = 0.165, 0.3367).The feedback-joystick did not yield a significant effect (g = 0.0467; p = 0.5191; 95% CI = -0.0952, 0.1886), which is probably a result of many studies using the

feedback-joystick in combination with implicit instructions. There was a significant difference between the feedback-joystick and the vertical stand (p = 0.0385) and between the

feedback-joystick and the joystick/lever (p = 0.0158). The moderator task explained 35% of the variance (tau² = 0.0111). The test for residual heterogeneity was significant (Qe =

38.6979, df = 21, p = 0.0107).

The test of the moderator instruction was significant (Qm = 17.6192, df = 2, p =

0.0001). The average effect size differed significantly from zero for explicit instructions (g = 0.2865; p < 0.0001; 95% CI = 0.204, 0.3689) and for incongruent instructions (g = 0.2874;

p < 0.0001; 95% CI = 0.1458, 0.429), but not for implicit instructions (g = 0.0281; p =

0.5716). The average effect size was significantly smaller for implicit instructions relative to explicit instructions (p < 0.0001) and relative to incongruent instructions (p = 0.0031). The moderator instructions explained 66% of the variance. The test for residual heterogeneity was not significant (Qe = 27.7543, df = 21, p = 0.1473).

Considering each level of the moderator stimulus type, the average effect size was significant for emotional words (g = 0.3391; p < 0.0001; 95% CI = 0.2109, 0.4673), as well as for emotional pictures (g = 0.2025; p = 0.0279; 95% CI = 0.0219, 0.3830), and for emotional facial expressions (g = 0.1482; p = 0.0017; 95% CI = 0.0559, 0.2406). Only the difference between emotional words and facial expressions was significant (p = 0.0179), which might also be due to many studies using facial expressions in combination with implicit instructions. However, the test of the moderator was not significant (Qm = 5.6186,

df = 2, p = 0.0602).

The test of the moderator design was not significant (Qm = 1.3386, df = 1, p =

0.2473), presumably due to the low number of studies having an independent groups design.

The test of the moderator valence was significant (Qm = 12.2724, df = 1, p =

significantly larger effect (g = 0.808; p < 0.0001; 95% CI = 0.4856, 1.1303) than all other studies (g = 0.2164; p < 0.0001; 95% CI = 0.1413, 0.2915). The moderator valence explained 45% of the variance. The test for residual heterogeneity was significant (Qe =

93.0283, df = 23, p < 0.0001).

Table 1

Results of Moderator Analyses of Positive Affect

Moderator Level k Estimate [95% CI] p p for diff

Task Ref Feedback 4 0.0467 [-0.0952, 0.1886] 0.5191 Stand 6 0.2721 [0.1127, 0.4316] 0.0008 0.0385 Stick 14 0.2509 [0.165, 0.3367] <0.0001 0.0158 Instruction Ref Implicit 7 0.0281 [-0.0693, 0.1256] 0.5716 Explicit 14 0.2865 [0.204, 0.3689] <0.0001 <0.0001 Incongruent 3 0.2874 [0.1458, 0.429] <0.0001 0.0031 Stimulus type Ref Faces 15 0.1482 [0.0559, 0.2406] 0.0017 Pictures 4 0.2025 [0.0219, 0.3830] 0.0279 0.6001 Words 5 0.3391 [0.2109, 0.4673] <0.0001 0.0179 Design Independent 1 0.6766 [-0.1069, 1.46] 0.0905 Repeated 23 0.212 [0.1368, 0.2872] <0.0001 0.2473 Valence Explicit 24 0.2164 [0.1413, 0.2915] <0.0001 Implicit 1 0.808 [0.4856, 1.1303] <0.0001 0.0005

Note. Ref = reference level for the comparison; k = number of studies; [95% CI] = 95%

confidence interval; p = p-value for each level; p for diff = p-value for difference between respective level and reference level. Task: Stand = vertical button stand, Joystick = joystick/lever, Feedback = feedback-joystick. Stimulus type: Words = emotional words, Pictures = emotional pictures, Faces = emotional facial expressions. Design: Repeated = repeated measures design, Independent = independent groups design. Valence: Explicit = explicitly valenced stimuli, Implicit = implicitly valenced stimuli.

Subset Analyses of Positive Affect

Because the moderator analyses above indicated no significant effect for implicit instructions, it is worthwhile to examine interactions between moderators. However, there needs to be at least one observation for each combination of moderator levels, which was not the case for the dataset of positive affect. Therefore, subset analyses were performed under exclusion of all studies using implicit instructions in order to investigate how they might affect the results.

Seven studies used implicit instructions. Under exclusion of these studies the random effects model yielded a significant medium average effect size (g = 0.2825; p < 0.0001; 95% CI = 0.2162, 0.3484). The estimated amount of heterogeneity was equal to

tau² = 0.0037; 95% CI = 0, 0.0211. The test for heterogeneity in effect sizes was not

significant (Q = 17.7367, df = 16, p = 0.3395), confirming that most of the heterogeneity was due to differences in average effect sizes between task-irrelevant instructions (implicit) and task-relevant instructions (explicit and incongruent).

As expected, moderator analyses for the subset of task-relevant instructions showed that no moderator was significant (except for the moderator valence). This

suggests that differences in the effect size between levels of the moderator task were due to an interaction with the moderator instructions. Specifically, the average effect size was

now significant for the feedback-joystick (g = 0.2334, p = 0.0353, 95% CI = 0.0160, 0.4508). The average effect size was not significantly different from the joystick/lever (p = 0.6895) and the vertical stand (p = 0.4628). The same applies to the moderator stimulus

type. There was no statistically significant difference in effect size between emotional facial

expressions and emotional words (p = 0.7066). Results are provided in Table 2.

Table 2

Results of Subset Analyses of Positive Affect

Moderator Level k Estimate [95% CI] p p for diff

Task Ref Feedback 1 0.2334 [0.0160, 0.4508] 0.0353 Stand 5 0.3365 [0.1679, 0.505] <0.0001 0.4628 Stick 11 0.2807 [0.1996, 0.3618] <0.0001 0.6895 Instruction Explicit 14 0.2863 [0.205, 0.3675] <0.0001 Incongruent 3 0.2843 [0.1472, 0.4215] <0.0001 0.9812 Stimulus type Ref Faces 8 0.2903 [0.1638, 0.4168] <0.0001 Pictures 4 0.199 [0.0450, 0.3530] 0.0113 0.3693 Words 5 0.3215 [0.2198, 0.4231] <0.0001 0.7066 Design Independent 1 0.6766 [-0.0732, 1.4263] 0.077 Repeated 16 0.2791 [0.2135, 0.3447] <0.0001 0.3006 Valence Explicit 17 0.2825 [0.2167, 0.3484] <0.0001 Implicit 1 0.808 [0.5797, 1.0362] <0.0001 <0.0001

Note. Ref = reference level for the comparison; k = number of studies; [95% CI] = 95%

confidence interval; p = p-value for each level; p for diff = p-value for difference between respective level and reference level. Task: Stand = vertical button stand, Joystick = joystick/lever, Feedback = feedback-joystick. Stimulus type: Words = emotional words, Pictures = emotional pictures, Faces = emotional facial expressions. Design: Repeated = repeated measures design, Independent = independent groups design. Valence: Explicit = explicitly valenced stimuli, Implicit = implicitly valenced stimuli.

Negative Affect

Figure 2 shows the funnel plot for the analysis of negative affect (k = 32). Four studies showed an effect in the opposite direction from the expected one. However, none of these effect sizes was significant. The remaining effect sizes were in the expected direction (k = 28). The effect sizes ranged from g = -0.13 to g = 1.85. Influence analysis identified two outliers, g = 1.85 (standardized residual = 1.5429) (Markman & Brendl, 2005; D), g = 1.76 (standardized residual = 1.4567) (Markman & Brendl, 2005; C).

Similar to the analysis of positive affect, the random effects model yielded a significant average effect size (g = 0.3042; p < 0.0001; 95% CI = 0.1735, 0.4350). The estimated amount of heterogeneity was equal to tau² = 0.1224; 95% CI = 0.0818, 0.3056. There was a clear indication of heterogeneity in effect sizes (Q = 189.969, df = 31, p < 0.0001). The exclusion of the two outliers resulted in a reduced average effect size (g = 0.2166; p < 0.0001; 95% CI = 0.1413, 0.2918). The unexplained variance component was diminished (tau² = 0.029), but the test for heterogeneity was still highly significant (Q = 104.3222, df = 29, p < 0.0001). Moderator analyses were conducted under exclusion of the two outliers.

Figure 2. Funnel plot of effect sizes for negative affect, representing a scatterplot of

treatment effect on the x-axis against the standard error on the y-axis (see Appendix B).

Moderator Analyses of Negative Affect

All five moderator variables (task, instruction, stimulus type, design, valence) were included in a mixed effects model. The results of the moderator analyses are provided in Table 3. The estimated amount of residual heterogeneity was equal to tau² = 0.0234. However, the test of the moderators was not significant (Qm = 13.7224, df = 9, p = 0.1325).

There was substantial residual heterogeneity (Qe = 56.8393, df = 20, p < 0.0001).

Separate analyses of each moderator confirmed that no moderator was significant, which means that none of the levels differed significantly from the other levels of the same moderator.

The test of the moderator task was not significant (Qm = 0.2646, df = 2, p = 0.8761).

The average effect size differed significantly from zero for the joystick/lever (g = 0.2351; p

< 0.0001; 95% CI = 0.1262, 0.344), as well as for the feedback-joystick (g = 0.2119; p =

0.0074; 95% CI = 0.0567, 0.367), and for the vertical stand (g = 0.1843; p = 0.0269; 95% CI = 0.0211, 0.3474).

The moderator instruction was close to significance (Qm = 5.9073, df = 2, p =

0.0521). Effect sizes differed significantly from zero for incongruent instructions (g = 0.389;

p = 0.0012; 95% CI = 0.1545, 0.6236) and for explicit instructions (g = 0.2493; p < 0.0001;

95% CI = 0.1594, 0.3392). Implicit instructions did not yield a significant effect (g = 0.1034;

p = 0.0959). Results indicated a significant difference between implicit and incongruent

instructions (p = 0.0341) and there was a trend towards a significant difference between implicit and explicit instructions (p = 0.0587). Further, there was clearly no difference between explicit and incongruent instructions (p = 0.2757). This pattern is consistent with the results of the analysis of positive affect. In order to gain power, the two levels of explicit and incongruent instructions were combined to form the level task-relevant instructions. The test of this moderator was significant (Qm = 4.7383, df = 1, p = 0.0295). Task-relevant

instructions showed a significant average effect size (g = 0.2672; p < 0.0001; 95% CI = 0.1834, 0.3510), whereas task-irrelevant (implicit) instructions did not (g = 0.1033; p = 0.0955; 95% CI = -0.0182, 0.2247).

The test of the moderator stimulus type was not significant (Qm = 5.5788, df = 3, p =

0.134). The effect size was significantly different from zero for personally relevant stimuli (g

= 0.3478; p = 0.0047; 95% CI = 0.1065, 0.5891), for emotional pictures (g = 0.3222; p =

0.0004; 95% CI = 0.1426, 0.5019), for emotional words (g = 0.2766; p = 0.0006; 95% CI = 0.1194, 0.4337), and for emotional facial expressions (g = 0.1344; p = 0.0089; 95% CI = 0.0336, 0.2352).

The test of the moderator design was not significant (Qm = 0.4934, df = 1, p =

0.4824).

One study used implicitly affective stimulus material (i.e., arrows). The effect size for this study was not significantly different from the effect size of all other studies (Qm =

Table 3

Results of Moderator Analyses of Negative Affect

Moderator Level k Estimate [95% CI] p p for diff

Task Ref Feedback 8 0.2119 [0.0567, 0.367] 0.0074 Stand 7 0.1843 [0.0211, 0.3474] 0.0269 0.81 Stick 15 0.2351 [0.1262, 0.344] <0.0001 0.81 Instruction Ref Implicit 10 0.1034 [-0.0183, 0.225] 0.0959 Explicit 17 0.2493 [0.1594, 0.3392] <0.0001 0.0587 Incongruent 3 0.389 [0.1545, 0.6236] 0.0012 0.0341 Stimulus type Ref Faces 16 0.1344 [0.0336, 0.2352] 0.0089 Pictures 5 0.3222 [0.1426, 0.5019] 0.0279 0.0739 Words 5 0.2766 [0.1194, 0.4337] 0.0006 0.1355 Relevant 4 0.3478 [0.1065, 0.5891] 0.0047 0.1098 Design Independent 1 0.5036 [-0.3007, 1.3078] 0.2197 Repeated 29 0.2141 [0.1383, 0.2898] <0.0001 0.4824 Valence Explicit 24 0.2028 [0.1295, 0.2761] <0.0001 Implicit 1 0.5182 [0.1589, 0.8775] 0.0047 0.0919

Note. Ref = reference level for the comparison; k = number of studies; [95% CI] = 95%

confidence interval; p = p-value for each level; p for diff = p-value for difference between respective level and reference level. Task: Stand = vertical button stand, Joystick = joystick/lever, Feedback = feedback-joystick. Stimulus type: Words = emotional words,

Pictures = emotional pictures, Faces = emotional facial expressions, Relevant = personally relevant stimuli. Design: Repeated = repeated measures design, Independent =

independent groups design. Valence: Explicit = explicitly valenced stimuli, Implicit = implicitly valenced stimuli.

Subset Analyses of Negative Affect

After recoding the moderator instruction, results showed a significant difference between task-relevant and implicit instructions. Although all other moderators were not significant, the magnitude of the average effect size for each level might depend on which instructions were used. Therefore, subset analyses were performed under exclusion of all studies using implicit instructions in order to investigate how they might affect the results.

Ten studies used implicit instructions. Under exclusion of these studies the random effects model yielded a significant medium average effect size (g = 0.2652; p < 0.0001; 95% CI = 0.1877, 0.3428). The estimated amount of heterogeneity was equal to tau² = 0.0178; 95% CI = 0.0058, 0.0618. The test for heterogeneity in effect sizes was significant (Q = 60.9215, df = 19, p < 0.0001).

Excluding studies using implicit instructions did not affect any of the omnibus moderator tests in terms of significance. As can be seen in Table 4, however, the magnitude of some average effect sizes was affected by the exclusion of implicit

instructions. Differences between levels of the moderator task were numerically reduced. Furthermore, one study using personally relevant stimuli showed a considerably larger effect size (g = 0.7692, p = 0.0058, 95% CI = 0.2224, 1.316).

Table 4

Results of Subset Analyses of Negative Affect

Moderator Level k Estimate [95% CI] p p for diff

Task Ref Feedback 2 0.244 [-0.0168, 0.5048] 0.0667 Stand 6 0.251 [0.0939, 0.4081] 0.0017 0.9642 Stick 12 0.2775 [0.1729, 0.3821] <0.0001 0.8153 Instruction Explicit 17 0.2476 [0.1646, 0.3306] <0.0001 Incongruent 3 0.3848 [0.1685, 0.6011] 0.0005 0.2457 Stimulus type Ref Faces 9 0.2020 [0.0761, 0.3280] 0.0017 Pictures 5 0.3171 [0.1580, 0.4762] <0.0001 0.2662 Words 5 0.2715 [0.1351, 0.4078] <0.0001 0.4633 Relevant 1 0.7692 [0.2224, 1.316] 0.0058 0.0476 Design Independent 1 0.5036 [-0.2733, 1.2805] 0.2039 Repeated 19 0.2629 [0.1847, 0.3411] <0.0001 0.5457 Valence Explicit 19 0.2465 [0.1725, 0.3204] <0.0001 Implicit 1 0.5182 [0.2273, 0.8091] 0.0005 0.076

Note. Ref = reference level for the comparison; k = number of studies; [95% CI] = 95%

confidence interval; p = p-value for each level; p for diff = p-value for difference between respective level and reference level. Task: Stand = vertical button stand, Joystick = joystick/lever, Feedback = feedback-joystick. Stimulus type: Words = emotional words, Pictures = emotional pictures, Faces = emotional facial expressions, Relevant = personally

relevant stimuli. Design: Repeated = repeated measures design, Independent = independent groups design. Valence: Explicit = explicitly valenced stimuli, Implicit = implicitly valenced stimuli.

Sensitivity Analyses

In order to investigate the impact of the correlation between pairs of observations, sampling variances were computed based on the two most extreme correlations. This was done separately for each analysis. For the analysis of positive affect, the lowest correlation derived from test-statistics was equal to r = 0.479 (Seidel, Habel, Kirschner, Gur, & Derntl, 2010; A). The highest correlation was equal to r = 0.797 (Dantzig, Pecher, & Zwaan, 2008; A). The results from the random effects model with the lowest correlation (g = 0.3104; 95% CI = 0.2008, 0.4199; Q = 181.1705; tau² = 0.0584) were very similar to those with the highest correlation (g = 0.2991; 95% CI = 0.1985, 0.3998; Q = 192.6461; tau² = 0.0546).

For the analysis of negative affect, the lowest correlation derived from test-statistics was equal to r = 0.403 (Rinck & Becker, 2007; Study 1). The highest correlation was equal to r = 0.966 (Dantzig, Pecher, & Zwaan, 2008; B). The results from the random effects model with the lowest correlation (g = 0.3128; 95% CI = 0.1724, 0.4533; Q = 166.7585;

tau² = 0.1297) were very similar to those with the highest correlation (g = 0.2998; 95% CI =

0.1696, 0.4301; Q = 440.7411; tau² = 0.1267).

In sum, these sensitivity analyses suggest that the effect size based on the weighted average of all available correlations are reliable and robust estimates.

Both Affects

Figure 3 shows the funnel plot for the analysis of both affetcs (k = 22). All effect sizes were in the expected direction. Eighteen of these effect sizes were significant. The effect sizes ranged from g = 0.002 to g = 0.87. Influence analysis identified no outliers. The

random effects model yielded a significant average effect size (g = 0.3076; p < 0.0001; 95% CI = 0.2052, 0.4099). The estimated amount of heterogeneity was equal to tau² = 0.045; 95% CI = 0.0213, 0.1107. The test for heterogeneity was significant (Q = 133.1284,

df = 21, p < 0.0001). The same five moderators were analyzed.

Figure 3. Funnel plot of effect sizes for both affects, representing a scatterplot of treatment effect on the x-axis against the standard error on the y-axis (see Appendix C).

Moderator Analyses of Both Affects

The results of the analyses are provided in Table 5. The estimated amount of residual heterogeneity was equal to tau² = 0.0007; 95% CI = 0, 0.0436, suggesting that 98.4% of the systematic variance in effect sizes could be accounted for by including the moderators (Qm = 104.4251, df = 9, p < 0.0001). The test for residual heterogeneity was

not significant (Qe = 15.8157, df = 12, p = 0.1998).

The test of the moderator task was not significant (Qm = 3.4415, df = 3, p =

0.3284).The average effect size differed significantly from zero for the vertical stand (g = 0.6617; p = 0.0039; 95% CI = 0.2124, 1.111), for the joystick/lever (g = 0.3047; p < 0.0001; 95% CI = 0.1865, 0.4228), and for the abstract task (g = 0.2811; p = 0.028; 95% CI =

0.0304, 0.5318). The feedback-joystick did not yield a significant effect (g = 0.0938; p = 0.6592).

The test of the moderator instruction was significant (Qm = 23.7076, df = 2, p <

0.0001). The average effect size was significantly different from zero for incongruent instructions (g = 0.433; p < 0.0001; 95% CI = 0.2948, 0.5712) and for explicit instructions (g = 0.4031; p < 0.0001; 95% CI = 0.2856, 0.5207). The difference between them was not significant (p = 0.7469). Implicit instructions did not yield a significant effect (g = 0.0764; p

= 0.1477). The average effect size was significantly smaller for implicit instructions than for

explicit instructions (p < 0.0001) and for incongruent instructions (p < 0.0001). The

moderator instruction explained 67% of the variance (tau² = 0.0150). The test for residual heterogeneity was significant (Qe = 50.5616, df = 21, p = 0.0001).

The test of the moderator stimulus type was not significant (Qm = 1.695, df = 2, p =

0.4285). The average effect size differed significantly from zero for emotional words (g = 0.3434; p < 0.0001; 95% CI = 0.2153, 0.4715), and for emotional pictures (g = 0.3206; p = 0.0119; 95% CI = 0.0708, 0.5705), but not for emotional facial expressions (g = 0.1462; p

= 0.2861).

The test of the moderator design was not significant (Qm = 2.4891, df = 1, p =

0.1146).

Two studies used implicitly affective stimulus material. The average effect size for these studies was not significantly different from the average effect size of all other studies (Qm = 0.4196, df = 1, p = 0.5171).

Table 5

Results of Moderator Analyses of Both Affects

Moderator Level k Estimate [95% CI] p p for diff

Task Ref Feedback 1 0.0938 [-0.3230, 0.5106] 0.6592 Abstract 3 0.2811 [0.0304, 0.5318] 0.028 0.4504 Stand 1 0.6617 [0.2124, 1.111] 0.0039 0.0693 Stick 17 0.3047 [0.1865, 0.4228] <0.0001 0.34 Instruction Ref Implicit 6 0.0764 [-0.0271, 0.18] 0.1447 Explicit 9 0.4031 [0.2856, 0.5207] <0.0001 <0.0001 Incongruent 7 0.433 [0.2948, 0.5712] <0.0001 <0.0001 Stimulus type Ref Faces 3 0.1462 [-0.1225, 0.4149] 0.2861 Pictures 4 0.3206 [0.0708, 0.5705] 0.0119 0.3515 Words 15 0.3434 [0.2153, 0.4715] <0.0001 0.1942 Design Independent 2 0.7492 [0.1898, 1.3052] 0.0086 Repeated 20 0.2912 [0.1895, 0.3929] <0.0001 0.1146 Valence Explicit 20 0.2966 [0.1874, 0.4059] <0.0001 Implicit 2 0.4087 [0.0876, 0.7298] 0.0126 0.5171

Note. Ref = reference level for the comparison; k = number of studies; [95% CI] = 95%

confidence interval; p = p-value for each level; p for diff = p-value for difference between respective level and reference level. Task: Stand = vertical button stand, Joystick = joystick/lever, Feedback = feedback-joystick, Abstract = abstract manikin task. Stimulus

type: Words = emotional words, Pictures = emotional pictures, Faces = emotional facial expressions, Design: Repeated = repeated measures design, Independent = independent groups design. Valence: Explicit = explicitly valenced stimuli, Implicit = implicitly valenced stimuli.

Subset Analyses of Both Affects

The moderator analyses so far have demonstrated consistently that task-relevant (explicit or incongruent) instructions are required to find an effect of affective information processing on approach-avoidance behaviors. Accordingly, the analysis of both affects also showed a non-significant effect for task-irrelevant (implicit) instructions. Again, subset analyses were performed under exclusion of all studies using implicit instructions in order to investigate how they might affect the results. However, inferences are based on even fewer studies and should therefore be treated with caution.

Six studies used implicit instructions. Under exclusion of these studies the random effects model yielded a significant medium average effect size (g = 0.4251; p < 0.0001; 95% CI = 0.3169, 0.5332). The estimated amount of heterogeneity was equal to tau² = 0.0283; 95% CI = 0.0071, 0.0877. The test for heterogeneity in effect sizes was significant (Q = 44.9764.9215, df = 15, p < 0.0001). Results are provided in Table 6.

The test of the moderator task was now slightly significant (Qm = 6.2184, df = 2, p =

0.0446). The level feedback-joystick was dropped, because the only study in this level used implicit instructions. There was only one study that used the abstract manikin task (g

= 0.7285, p < 0.0001, 95% CI = 0.3727, 1.0844). This effect size was almost significantly

larger than for the joystick/lever (g = 0.3678, p < 0.0001, 95% CI = 0.2681, 0.4675) (p = 0.0557). The moderator task explained 42% of the variance (tau² = 0.0164). The test for residual heterogeneity was significant (Qe = 27.0039, df = 13, p = 0.0124).

0.2404). Based on two studies, the average effect size for emotional facial expressions was still not significant (g = 0.2098, p = 0.1376, 95% CI = 0.0672, 0.4867).

Table 6

Results of Subset Analyses of Both Affects

Moderator Level k Estimate [95% CI] p p for diff

Task Ref Abstract 1 0.7285 [0.3727, 1.0844] <0.0001 Stand 1 0.6617 [0.3481, 0.9753] <0.0001 0.7823 Stick 14 0.3678 [0.2681, 0.4675] <0.0001 0.0557 Instruction Explicit 9 0.4147 [0.2674, 0.562] <0.0001 Incongruent 7 0.4426 [0.271, 0.6142] <0.0001 0.809 Stimulus type Ref Faces 2 0.2098 [-0.0672, 0.4867] 0.1376 Pictures 3 0.4229 [0.1757, 0.6702] 0.0008 0.2604 Words 11 0.4740 [0.3420, 0.6060] <0.0001 0.0913 Design Independent 2 0.7492 [0.2187, 1.2796] 0.0056 Repeated 14 0.4108 [0.3016, 0.5199] <0.0001 0.2207 Valence Explicit 15 0.4312 [0.3075, 0.5548] <0.0001 Implicit 2 0.4062 [0.1303, 0.6821] 0.0039 0.8716

Note. Ref = reference level for the comparison; k = number of studies; [95% CI] = 95%

confidence interval; p = p-value for each level; p for diff = p-value for difference between respective level and reference level. Task: Stand = vertical button stand, Joystick =

joystick/lever, Abstract = abstract manikin task. Stimulus type: Words = emotional words, Pictures = emotional pictures, Faces = emotional facial expressions. Design: Repeated = repeated measures design, Independent = independent groups design. Valence: Explicit = explicitly valenced stimuli, Implicit = implicitly valenced stimuli.

Publication Bias

Visual inspection of the funnel plots (Figure 1 through Figure 3) suggests that all plots are asymmetrical. This is confirmed by a significant regression test for funnel plot asymmetry for all three analyses. The regression test suggested asymmetry in the funnel plot for positive affect (p = 0.0233). The asymmetry, however, was dependent on the three outlying cases. After excluding them, the test was not significant (p = 0.0507). The same applied to the funnel plot for negative affect. There was a statistical sign of asymmetry (p = 0.0002), which vanished after excluding the two outlying cases (p = 0.1088). The test also suggested asymmetry in the funnel plot for both affects (p < 0.0001), indicating that a publication bias was present. Thus, the tests all give results consistent with some degree of publication bias. According to the trim-and-fill method, however, only the funnel plot for both affects showed substantial asymmetry. As shown in Figure 4, the estimated number of missing studies on the left side was 5. After adjusting for the missing studies, the estimated overall effect size dropped from g = 0.3076 to g = 0.2361.

Figure 4. Funnel plot of effect sizes for both affects, representing a scatterplot of treatment effect on

the x-axis against the standard error on the y-axis. Diamonds are original data (see Appendix C), open circles represent filled-in data based on the trim-and-fill method.

Discussion

The aim of the present meta-analysis was to systematically review the literature in order to assess the magnitude of the effect of affective evaluation on approach-avoidance behaviors. Separate analyses were performed for (1) stimuli with a positive affective valence, (2) stimuli with a negative affective valence, and (3) for a combination of both positive and negative stimuli. In addition, several potential moderators were included in order to examine the automaticity of the link between affect and action-tendencies of approach and avoidance and to investigate existing accounts of compatibility relations between affective stimuli and arm movements.

The results of the meta-analysis showed a significant overall effect in studies

conducted between 1999 and 2012. We found significant medium-sized effects for all three analyses, ranging from 0.22 to 0.31. There was, however, an indication of a small

and the correction for publication bias according to the trim-and-fill procedure resulted in a reduced effect size of 0.24. It can therefore be concluded that positive and negative affect have similarly sized effects on the approach-avoidance task.

A consistent finding across all analyses was a non-significant overall effect when instructions did not require conscious appraisals of the affective valence of stimuli. In these studies there could only be an implicit relation between affect and approach and avoidance. Our results showed no effect on approach-avoidance movements when subjects were instructed to evaluate a feature of the presented stimulus other than its affective valence (i.e., task-irrelevant instructions). This thus seems especially true for tasks involving arm movements and is in line with the conclusion based on the study by Rotteveel and Phaf (2004). The vertical button stand may be less liable to automatic associations between affect and approach and avoidance movements than the other tasks (cf. Alexopoulos & Ric, 2007). The implicit studies using the joystick/lever, however, yielded a similar small effect size, which may indicate that, in general there does not seem to be an automatic link between affective information processing and arm flexion and extension. Instead, this link depends on intentions to respond to affectively laden stimuli.

This conclusion can, however, not be generalized to all kinds of approach and avoidance tasks and experimental designs. When considering single studies included in the meta-analysis, reliable effects of task-irrelevant instructions could be observed for certain experimental setups. For instance, Najmi, Kuckertz, and Amir (2010) used the feedback-joystick task to assess approach-avoidance tendencies in individuals with contamination-related obsessive-compulsive symptoms. They found rather large effects, although subjects were instructed to respond to the orientation of stimuli. As mentioned earlier, however, compatibility effects obtained with the feedback-joystick task might be solely due to the zooming feature of the task. The zooming feature has been shown to be resistant to cognitive re-interpretations. In their second experiment (not included in the

current meta-analysis), Rinck and Becker (2007) phrased the instructions so that pulling the joystick was described as pulling it away from the stimulus (i.e., avoidance), and pushing the joystick was described as pushing it toward the stimulus (i.e., approach), also referred to as an object-related frame of reference. However, pulling the joystick still increased, whereas pushing the joystick decreased the size of the stimulus. They still found a compatibility effect. That is, pulling away from positive stimuli was faster than pushing toward positive stimuli and pushing toward negative stimuli was faster than pulling away from negative stimuli. This effect was not smaller than the effect they found with a self-related frame of reference in their third experiment (i.e., pulling the joystick toward the self vs. pushing it away from the self), where an increase and decrease in stimulus size corresponded to approach and avoidance, respectively. Thus, instead of the arm

movements, it is most likely the zooming function, which drives the compatibility effect observed with the feedback-joystick task.

As far as irrelevant instructions are concerned, the results from the present meta-analysis resemble those of the study conducted by Krieglmeyer and Deutsch (2010). They compared different measures of approach-avoidance behavior. When subjects were

instructed to respond to the grammatical category of emotional words (task-irrelevant instructions), only the manikin task and the feedback-joystick task but not the joystick task were sensitive to valence. In the manikin task, an abstract manikin appears on the screen and is controlled by simple button presses. Therefore, this task does not involve any arm movements at all, supporting the idea that some approach-avoidance behaviors do not depend on conscious appraisals.

Our results also showed that the effect was not moderated by the type of stimuli used. All stimuli yielded a significant effect on approach-avoidance behaviors, but there was no difference between them. At first sight this seems to be at odds with the idea of emotional facial expressions being evolutionary prepared stimuli (Öhman, 1986), which

receive processing priority. If anything, there was a tendency for facial expressions to be less effective in initiating approach-avoidance behavior than all other stimuli. On the other hand, the fact that in the present meta-analysis, emotional facial expressions did not show to be a special category of stimuli, can be interpreted as further evidence that affective information processing does not automatically predisposes approach-avoidance behavior. In their third experiment, Rotteveel and Phaf (2004) presented facial expressions as primes prior to affective scenes, to which subjects were instructed to respond by flexing or extending the arm. If there is an automatic link to action-tendencies, the prime faces should influence arm flexion and extension reaction times, just as emotional facial expressions should yield larger effects than other stimuli. They did not find an effect on arm flexion and extension reaction times. Importantly, however, they did find a priming effect in the overall responses to the affective scenes. In other words, affective information processing might occur automatically, but the follow-up link with arm movements is

dependent on intentions to respond to the prime faces. This conclusion is supported by the finding that the type of stimuli does not moderate the link between affect and approach-avoidance arm movements.

Thus far, it has been established that the link between affect and approach-avoidance behavior, defined as arm flexion and extension, is not entirely automatic but depends on intentions to respond to affective stimuli. A second aim of this meta-analysis was to examine whether this link can be accounted for in terms of fixed effects of biceps and triceps activations. Is the effect of arm flexion and extension inherently and inflexibly related to approach and avoidance, respectively, or does the meaning of these movements vary depending on the context in which they are embedded? This question can be

addressed by comparing different approach-avoidance tasks and examine the exact manner in which task-relevant instructions are phrased.

the effect differed between tasks. Obviously, there are conceptual similarities between arm flexion and extension as initiated with the vertical three-button stand and pulling and pushing of a joystick or a lever. Our results suggest that the vertical stand and the joystick measure similar conceptual mechanisms. There are, however, theoretical reasons to regard the vertical stand as a purer measure approach-avoidance behavior in terms of arm flexion and extension. Arm movements are made in the vertical direction by either flexing the arm with the biceps muscle or extending the arm with the triceps muscle. Instructions only refer to the upper and lower button with no explicit references to approach and avoidance behaviors.

In contrast, pulling and pushing movements of a joystick or lever also involve pulse and shoulder muscles. Different accounts can explain compatibility effects measured with a joystick or lever. Specifically, both fixed effects of arm flexion and extension and

accounts in terms of cognitive labeling of these actions can explain compatibility effects. We tried to validate the plausibility of these different accounts by comparing different instructions. Instructions that induced a self-related frame of reference were here referred to as explicit instructions. With explicit instructions, subjects are encouraged to interpret a pulling movement as approach (i.e., move the joystick toward the self) and a pushing movement as avoidance (i.e., move the joystick away from the self). We found a significant effect of explicit instructions that was not significantly different from what we referred to as incongruent instructions. Incongruent instructions reverse the coupling between valence and the flexor and extensor movements. This is usually achieved by inducing an object-related frame of reference, as was done, for instance, in the study by Lavender and

Hommel (2007). In this case, interpretations of arm movements are exactly opposite to the presumed fixed effects in terms of approach and avoidance. Flexing the arm now

corresponds to avoidance, whereas extending the arm reflects an approach movement. Eder, Rothermund, and Proctor (2010) suggested that the direction of the effect depends

on how participants construe the flexor and extensor movements, which may depend on their intentions on how to deal with the affective stimuli (i.e., implementation intentions). Such intentions do not always involve conscious planning but may be set up by

unobtrusive manipulations or even inadvertently by specific features of the design. In this view, the processing of affect can still be automatic but the coupling of affect to specific movements depends on the interpretations of the movement. In the studies using implicit instructions and implicit presentation conditions, no intentions may have been set up at all. In fact, this reasoning can also explain the relatively large effect found in the study by Phaf and Rotteveel (2009). The affective monitoring framework (Phaf & Rotteveel, 2012) argues that a correspondence between arrow direction and habitual eye movements made in the reading direction elicits positive affect. Although subjects were not spontaneously aware of the affective valence of the arrows, implementation intentions may have been raised by the context of the other experiments that were conducted along with the arrow experiment, which all involved explicit evaluation of color stimuli.

Our finding of a significant effect for explicit as well as incongruent instructions can be taken as evidence that there are default values of arm flexion and extension that are superimposed by their situated meaning implied by incongruent instructions. With an object-related frame of reference, arm movements might then be coded in terms of distance regulations. Eder and Rothermund (2008) even found compatibility effects when joystick movements to the right and left were referred to as toward and away movements, respectively. In sum, the effect of incongruent instructions casts serious doubt on the notion of fixed effects of arm flexion and extension. In fact, we even found a trend for incongruent instructions to yield larger effects than explicit instructions, suggesting that default values of arm movements are either non-existent or simply replaced by contextual re-interpretations. Thus, an evaluative-response-coding account (Eder & Rothermund, 2008) seems to be able to integrate compatibility effects found in this meta-analysis.

According to this view, approach and avoidance behaviors are not regulated by distinct motivational mechanisms. Evaluative codes are assigned to arm movements and

compatibility effects are due to a match between these codes and the valence of stimuli. As far as a joystick or lever is concerned, an explanation of the link between

affective information processing and approach-avoidance behaviors in terms of evaluative codes assigned to arm movements seems plausible. This explanation, however, is unlikely to apply to the effects found with the vertical three-button stand. As mentioned earlier, instructions are not contaminated with explicit references to approach and avoidance. Subjects are simply told to press the upper or lower button. Eder and Rothermund (2008) also found compatibility effects when movements were labeled as upwards and

downwards. Experiments with the button stand, however, were mostly conducted in the Netherlands. The terms used in the Dutch language to phrase the instructions are not as clearly positively or negatively connoted as their English counterparts. Furthermore, compatibility effects cannot be explained in terms of a distance-regulation account, because movements are made in the vertical direction, holding the distance constant between the self and an object.

In summary, our meta-analysis has established a small to medium-sized effect of affective evaluation on approach-avoidance behaviors. This effect requires affective

interpretations of the movements and implementation intentions with regard to the affective stimuli. Furthermore, our review suggests that there are default effects of arm flexion and extension in terms or approach and avoidance, respectively. These meanings of arm movements, however, seem to be easily replaced by alternative interpretations, which are more effective in specific contexts.

Acknowledgements

We thank Dr. Eva-Maria Seidel, Dr. Karin Roelofs, Dr. Matthew Field, Dr. Karin Mogg, Dr. Reinecke, Dr. Mike Rinck, and Dr. Cigdem Önal-Hartmann for providing additional data from their publications.

References

*References marked with an asterisk indicate studies included in the meta-analysis.

Alexopoulos, T., & Ric, F. (2007). The evaluation-behavior link: Direct and beyond valence.

Journal of Experimental Social Psychology, 43, 1010-1016.

* Bakvis, P., Spinhoven, P., Zitman, F. G., & Roelofs, K. (2011). Automatic avoidance tendencies in patients with psychogenic non epileptic seizures. Seizure, 20, 628-634.

Cacioppo, J. T., Priester, J. R., & Berntson, G. G. (1993). Rudimentary determinants of attitudes. II: Arm flexion and extension have differential effects on attitudes. Journal

of Personality and Social Psychology, 65, 5-17.

Centre for the Study of Emotion and Attention [CSEA-NIMH] (1995). The international

affective picture system: Photographic slides. Gainsville, FL: The Center for

Research in Psychophysiology, University of Florida.

* Chen, M., & Bargh, J. A. (1999). Immediate behavioral predispositions to approach or avoid the stimulus. Personality and Social Psychology Bulletin, 25, 215-224. Clow, K. A., & Olson, J. M. (2010). Conceptual-motor compatibility and homonegativity:

approaching and avoiding words associated with homosexuality. Canadian

Journal of Behavioural Science, 42, 222-233.

* De Houwer, J., Crombez, G., Baeyens, F., & Hermans, D. (2001). On the generality of the affective Simon effect. Cognition and Emotion, 15, 189-206.