Frontal asymmetry as a mediator and moderator of emotion: An updated review

1.1 A brief summary of what rLFA measures

Frontal EEG asymmetry is a measure of the difference in EEG alpha power between homologous right and left frontal electrodes. Since alpha activity has an inhibitory influence on cortical network activity, lower frontal asymmetry scores (right minus left alpha) putatively reflect less left than right cortical activity (Allen, Coan, & Nazarian, 2004). Alpha activity is typically measured at 8–13 Hz, and frontal asymmetry is calculated as the natural log of the right minus left hemisphere (ln[right] − ln[left].) Activity is typically investigated at midfrontal (F4-F3) and lateral frontal (F6-F5, F8-F7) sites. Frontal pole (Fp2-Fp1) sites are less commonly examined. For research involving infants, the alpha band is often considered to be lower, around 6–9 Hz (Smith & Bell, 2010).

As a relative measure, lower frontal asymmetry scores reflect either less left than right cortical activity or more right than left cortical activity. For purposes of brevity and clarity, frontal asymmetry scores in this review will be discussed in terms of relative left frontal activity (rLFA). As such, less rLFA reflects lower frontal asymmetry scores and more rLFA reflects higher frontal asymmetry scores. Additionally, differences between groups or conditions can result from significant differences at either, neither, or both of constituent recording electrodes. As such, activity at each hemisphere must be investigated independently in order to determine which hemisphere contributes to changes in the difference score (Allen, Coan, & Nazarian, 2004; Harmon-Jones & Allen, 1997).

Most studies examine frontal asymmetry either at rest and as a trait measure of psychological phenomena (Allen & Cohen, 2010; Peltola et al., 2014; Stewart, Bismark, Towers, Coan, & Allen, 2010; Thibodeau, Jorgensen, & Kim, 2006) or during emotionally evocative tasks as a state measure of current emotion or behavior (Harmon-Jones, 2007; Killeen & Teti, 2012; Stewart, Coan, Towers, & Allen, 2011). Consequently, the literature generally focuses on two different constructs: frontal EEG “activity” and “activation.” Activity refers to recording cortical activity at a given time (e.g., at rest, or while watching a film), whereas activation refers to a change in EEG activity in response to stimuli or task (e.g., the difference from rest to the emotion). Further complicating this distinction, task-elicited rLFA is sometimes assessed as activity during the task and at other times activation reflecting change from a nontask period.

The approach-withdrawal motivational model of frontal asymmetry proposes a lateralization of affect across the right and left frontal hemispheres. Left frontal regions are primarily involved with appetitive motivation and approach-related affect such as elation, hope, happiness, and anger (Depue & Iacono, 1989; Harmon-Jones, 2003). In contrast, right frontal regions are associated with behavioral inhibition and vigilant attention that often occurs during certain negative affective states (Coan & Allen, 2004; Davidson & Irwin, 1999; Harmon-Jones & Allen, 1998). The approach system is responsible for goal-directed behavior and appetitive motivation to reward stimuli, whereas the withdrawal system responds to punishment and the termination of reward (Davidson & Irwin, 1999).

Evidence that rLFA is related to behavioral activation (BAS) and behavioral inhibition (BIS), as measured by Carver and White's (1994) scale, supports this model. A number of studies have found that rLFA is related to BAS (Alloy et al., 2008; Coan & Allen, 2003a; De Pascalis, Cozzuto, Caprara, & Alessandri, 2013; Harmon-Jones & Allen, 1997). Yet, not all studies have found that rLFA relates to BAS or agentic extraversion (Wacker, Chavanon, & Stemmler, 2010); possibly due to construct heterogeneity related to the original conceptualization of approach and withdrawal proposed by Davidson (1992). BAS involves the approach-related processes having to do with the attainment (or failed attainment) of rewards or goals. BIS, by contrast, inhibits prepotent approach or avoidant behaviors when there may be competing motivations (Gray & McNaughton, 2000). When assessed with self-report (Carver & White, 1994), BIS may more specifically reflect sensitivity to punishment and threat (Alloy, Abramson, Urosevic, Bender, & Wagner, 2010), thus failing to capture the conflict management functions of the BIS (Wacker, Chavanon, Leue, & Stemmler, 2008).

The isomorphic mapping of BIS and BAS systems to withdrawal and approach systems has been challenged (Harmon-Jones & Allen, 1997). Withdrawal defined within Davidson's model (Davidson & Irwin, 1999) includes all behavior that relates to withdrawing from stimuli but BIS involves interrupting behavior, which does not preclude increased attention or approach-oriented behavior. Additionally, BAS involves the monitoring of reward stimuli, whereas approach motivation involves appetitive behavior. As such, BAS could motivate seeking safety cues elsewhere, which could be considered a withdrawal behavior. See Coan & Allen (2003a) for a more extensive considerations of how BAS/BIS may relate to approach/withdrawal constructs.

An additional challenge to note here is that, although a number of studies reveal correlations between self-reported BAS sensitivity and rLFA, the correlations are of an insufficient magnitude to suggest that rLFA is a synonymous construct with BAS sensitivity.

1.2 Methodological issues in frontal asymmetry research

Although a substantial body of literature has linked rLFA to approach and approach-related disorders, differences in methodology between studies have led to some inconsistent findings. rLFA scores may be impacted by, among other factors, the type of task used (i.e., resting or emotionally evocative and the extent to which cognition is involved), recording montage, and site utilized in the analyses.

First, experimental conditions may influence the measurement of frontal asymmetry in potentially substantial ways. For instance, the capability model proposes that frontal asymmetry may be an interaction between the emotional salience of a situation and an individual's emotion predisposition (Coan, Allen, & McKnight, 2006). Therefore, individual differences in frontal asymmetry may be most pronounced during an emotionally evocative task rather than at rest. In support of this model, Stewart, Coan, and colleagues (2011) found that frontal asymmetry was more strongly related to depression status during a directed facial action task than at rest (for data referenced to Cz, linked mastoids, and the averaged reference, but not for current-source-density-transformed data for which rLFA was related to depression status under both conditions).

Second, the wide range of tasks and stimuli used often involve many different emotional and cognitive processes. It is likely that cognitive processes significantly impact rLFA (Papousek & Schulter, 2004) and individuals with less rLFA may have more impaired reasoning in depression (Brzezicka, Kamiński, Kamińska, Wołyńczyk-Gmaj, & Sedek, 2016). For example, Meyer et al. (2014) found a significant difference in response to two trauma-related films that differed in their difficulty of cognitive reappraisal. Therefore, it may be important to consider the degree to which a task constrains cognition. Resting state, for example, is the least constrained task used in rLFA research. During resting state, each individual will engage in a completely individualized process that will involve both cognitive and affective components that have not been constrained by the researcher. As such, there may be greater moment-to-moment variability in resting asymmetry than during task-related asymmetry in which both cognitive and affective processing may be more constrained.

Additionally, individual differences in cognitive ability may moderate rLFA response to emotional tasks. For example, Papousek et al. (2016) found that individuals who have higher ability to cognitively reappraise potentially anger-provoking material are more likely to demonstrate more rLFA during a cognitive reappraisal task. Although moderators of rLFA and other variables will not be discussed in the present paper, it is important to note such a substantial moderator.

Gable and Harmon-Jones (2008) also found that the personal relevance of the stimuli might impact rLFA findings. In this study, participants indicated their liking of dessert before being shown dessert and neutral pictures as EEG was recorded. Participants who liked dessert more had more rLFA in response to dessert pictures. In another study, participants with greater opposition to racism had more rLFA in response to pictures depicting racism (Harmon-Jones, Lueck, Fearn, & Harmon-Jones, 2013). Steiner and Coan (2011) also examined the relationship between rLFA in both experiential reports of homesickness in college freshman and retrospective reports of freshman year homesickness. They found that rLFA predicted only the experiential but not retrospective reports, suggesting that rLFA may relate to current emotional challenge but not memory of affective state over time.

Recording montage differences between studies may lead to substantial variability across studies and thus possible inconsistencies in results. The most commonly used recording montages are an average reference (average of activity at each site), CZ (activity at vertex), and linked mastoids (activity at mastoid bones averaged) reference. The CSD transformation may be a more sensitive measure of variations in frontal asymmetry per se, rather than alpha activity originating elsewhere. CSD uses an estimate of sources and sinks on the scalp in order to minimize any effects of distal sources. If measures of frontal asymmetry are to be interpreted as a likely reflection of activity in frontal regions, then the CSD transform is preferred, as traditional reference montages may be more susceptible to volume-conducted activity. In the largest study of depressive risk and EEG asymmetry to date (n = 306), only current-source-density (CSD) transformed frontal asymmetry—but not average, linked, mastoids, nor CZ referenced frontal asymmetry—distinguished between individuals with and without a lifetime history of depression (Stewart, Coan et al., 2011). As there is only a small correlation between CSD-transformed data and data measured with traditional reference montage (Hagemann, Naumann, & Thayer, 2001), it is possible that well-established relationships between frontal asymmetry and other variables may fail to emerge when studied with CSD-transformed frontal asymmetry. Such a finding would suggest that such effects attributed to frontal asymmetry might reflect asymmetric volume-conducted activity from nonfrontal regions, an important interpretive consideration. We suggest that future research investigate reference as a moderator of frontal asymmetry relationships.

Further consideration of site and the potential differences between midfrontal and lateral frontal regions may be helpful in clarifying the role of rFLA in psychopathology. Some research has found that rLFA is related to psychopathology only at midfrontal regions (Minnix & Kline, 2004; Pössel, Lo, Fritz, & Seemann, 2008), whereas others have found relationships between rLFA and emotional responding at lateral frontal sites (Amodio, 2010). Research typically examines these sites independently, and the field lacks an a priori theory as to why different regions may be involved. Additionally, as the term rLFA encompasses both midfrontal, lateral frontal, and sometimes even frontal pole regions, researchers must take care to explicitly state the sites involved in each particular analysis. Since activity at multiple sites will be less highly correlated using CSD-transformation, CSD may have greater signal specificity. This more precise measurement at midfrontal versus lateral-frontal sites may lead to theoretical models that begin to subdivide the somewhat vague concept of “frontal” asymmetry into a more regionally specific function.

Finally, other often unmeasured factors may influence rLFA. Such effects include the time of day or year (Peterson & Harmon-Jones, 2009; Velo, Stewart, Hasler, Towers, & Allen, 2012) as well as aspects of the experimental interaction (e.g., Blackhart, Kline, Donohue, LaRowe, & Joiner, 2002; Wacker, Mueller, Pizzagalli, Hennig, & Stemmler, 2013)

More detailed reviews of methodological considerations and processing of asymmetry data are available. The reader is referred to Allen, Coan, & Nazarian (2004) and Smith, Reznik, Stewart, & Allen (2016).

1.3 rLFA as a predictor and outcome variable

Cacioppo (2004) draws an important distinction between studies in which rLFA serves as the outcome and studies in which rLFA serves as the predictor variable. When rLFA serves as the predictor variable, the results indicate the probability of a given correlate (behavioral measure) given rLFA. On the other hand, when rLFA serves as the outcome variable, the results describe the probability of rLFA given a particular behavioral measure. Some studies make no distinction and treat rLFA and behavioral measures as simple correlates.

1.4 Frontal asymmetry as a moderator of emotion

A moderator is a third variable that alters the relationship between the predictor and the outcome (Baron & Kenny, 1986). Most commonly, moderators emerge when the relationship between two variables is different for different groups of people or for those with different scores on a continuous measure of individual differences. rLFA is a moderator when some relationship differs as a function of whether one has a high or low rLFA score. For example, Bruder et al (2001) found that those who responded to antidepressant treatment had more rLFA at rest. Both rLFA (frontal asymmetry activity) and change in rLFA (frontal asymmetry activation) may serve as potential moderators, as described below (Dennis & Solomon, 2010).

Moderators can be enhancing, buffering, or antagonistic. When a moderator is enhancing, a higher level of the moderator variable increases the effect of the predictor on the outcome. When a moderator is buffering, a higher level of the moderator variable decreases the effect of the predictor on the outcome. When a moderator is antagonistic, the effect of the predictor on the outcome is reversed at different levels of the moderator variable (Breitborde, Srihari, Pollard, Addington, & Woods, 2010). Figure 1 illustrates the conceptual relationship between a moderator and the predictor and outcome variable.

Statistically, moderators emerge as interactions between the third variable and the predictor variable (see Figure 1). In other words, the effect of one variable on another depends upon the level of a third variable. Main effects of the predictor on the outcome and the third variable on the outcome may or may not be present (MacKinnon, Lockhart, Baraldi, & Gelfand, 2013). In fact, moderators often emerge when it appears that there is no relationship between the predictor and the outcome. For example, imagine that a researcher does not find a relationship between a therapeutic intervention and self-reported positive mood. It is possible that treatment does not impact any group. However, it may also be possible that the treatment greatly improved mood for people with more rLFA and did not change mood for people with less rLFA. In this hypothetical example, an important relationship emerges because of the addition of a moderator. However, it is important that the addition of moderators is theoretically driven. Without theory, it may be difficult to infer whether a moderator is spurious. In this particular example, one could suppose that those with more rLFA had greater approach motivation and as such had higher motivation and engagement with the intervention. It would be important to examine that both those with more and less rLFA had equivalent mood symptoms at the outset of the study for this theory to be pragmatic. As such, consideration of theoretically driven potential moderators at the outset of a study may provide significant insight into for whom a treatment works or under what conditions a particular relationship exists.

Moderators should be distinguished from covariates, which are third variables statistically related to the outcome variable. Adjusting for a covariate clarifies the relationship between the predictor and the outcome by reducing the variance attributed to the relationship between the predictor and the outcome. As such, there is an important conceptual distinction between moderators and covariates in that moderators change the relationship, whereas covariates clarify it by adjusting for another variable (see Smith, Reznik et al., 2016 for illustration of the conceptual distinction between moderator and covariate). In the example above, the relationship between treatment and outcome changed based on whether one had less or more rLFA. If there was no interaction between rLFA and the predictor on the outcome but rather rLFA was related only to the outcome, it may be reasonable to adjust for rLFA as a covariate. True covariates explain variance in the outcome but not predictor variables. If a third variable explains variance in both the outcome and predictor, including such a variable as a covariate may lead to ambiguous results (Kraemer, 2015; Miller & Chapman, 2001). For example, if examining the effect of a treatment on depressive symptoms, rLFA may only be treated as a covariate in the case that it relates to the depressive symptoms but not the treatment itself. Statistically, covariates may be identified when there is a statistically significant main effect of the third variable but no interaction between the predictor and third variable on the main effect. Moderators, on the other hand, must have a significant interaction between them and the predictor variable.

1.5 Evidence supporting frontal EEG asymmetry as moderator of emotion

Although the vast majority of frontal asymmetry research examines rLFA as a predictor or outcome, there is some evidence that suggests rLFA may serve as a moderator of emotional response in both infants and adults. See Table 3 for a summary of rLFA as a moderator of emotion. Additionally, there is some evidence that rLFA in infancy may moderate the relationship between emotional responses and behavior. Infants with less rLFA are more likely to cry in response to separation from their mothers (Davidson & Fox, 1989; Fox, Bell, & Jones, 1992). Infants with less rLFA at 9 months had negative emotionality scores that predicted their social wariness 4 years later. On the other hand, this predictive relationship did not exist in infants with more rLFA (Henderson, Fox, & Rubin, 2001).

There is also some evidence that rLFA may serve as a moderator of the relationship between infant and mother affect. Theoretically, this research suggests that both less rLFA and having a mother with depression represent risk factors for developing depression. However, one is most likely to develop depression if one has both low rLFA and a mother with depression. Forbes et al. (2008) examined the relationship between depressive symptoms in mothers and negative affect in their 3- to 9-year-old children. They found that for children with less rLFA at rest, mothers' depressive symptoms predicted negative affect in their children. This finding suggests that rLFA is a moderator of the relationship between mother and child's depression; that is, a mother's depressive symptoms predict a child's depressive symptoms only for those children who have less rLFA, with greater rLFA indexing a potential buffering mechanism between maternal depression and negative affect in their children.

Swingler, Perry, Calkins, and Ann (2014) examined rLFA at rest in 233 5-month-old infants as a moderator of the relationship between maternal sensitivity and infant response to an emotional challenge task. During the emotional challenge task, mothers were asked to hold their infants' arms by their sides to restrict their movement, while maintaining a neutral facial expression. Infants' regulatory behavior (i.e., orienting to mother or distracting attention to other things in the room) and affect (as rated by observer) were recorded. For infants with more rLFA at rest, maternal sensitivity predicted distraction during the emotional challenge task. For infants with less rLFA, maternal sensitivity predicted negative affect during the emotional challenge. These results suggest that infant rLFA may serve as a moderator between maternal sensitivity and infant emotional response. These findings suggest that rLFA may play a role in moderating emotional response or the relationship between emotional response and other trait variables such as maternal sensitivity. Similar findings exist in the adult literature.

In particular, Tomarken, Davidson, and Henriques (1990) found that those with less rLFA at rest had greater self-reported negative affect in response to negative emotional film clips than those with more rLFA. As such, rLFA serves as a moderator of the relationship between film clips and self-reported emotional response, such that those with less rLFA report more negative affect in response to negative stimuli. A subsequent study found that those with more rLFA had greater self-reported positive affect in response to positive emotional films than those with less rLFA (Wheeler, Davidson, & Tomarken, 1993). Recently, Schmidt and Hanslmayr (2009) examined ratings of mood and enjoyment in response to musical stimuli. They found that participants with more rLFA rated the musical stimuli more positively than those with less rLFA. Additionally, self-reported enjoyment of the stimuli differed as a function of rLFA for negative musical stimuli. This work suggests that rLFA at rest changes the relationship between the emotional stimuli (predictor) and self-rated affect and enjoyment (outcome). More generally, these studies indicate that frontal asymmetry is implicated as a potential moderator of emotional responding for both approach- and withdrawal-related emotions. Having less rLFA predicts increased attention to negative affect in response to withdrawal-related stimuli, which may be less adaptive and pose a potential risk factor for disorders characterized by deficits in approach-related motivation and affect, such as depression. More rLFA may predict increased positive affect to approach-related stimuli and indicate less vulnerability to such disorders.

Gatzke-Kopp, Jetha, and Segalowitz (2014) examined rLFA as a potential moderator of emotional response to sad stimuli and internalizing and externalizing symptoms in kindergarten children. This study was theory-driven in exploring whether rLFA served as an “afferent” or “efferent” filter. The afferent model proposes that rLFA is a filter of emotional stimuli that influences the behavioral response via affective arousal. The efferent model of rLFA proposes that rLFA alters the behavioral response to affective arousal without changing affective arousal. They found that rLFA served as a moderator; those who had increased arousal in response to the sad clip with more rLFA had more externalizing symptoms, whereas those with less rLFA had more internalizing symptoms. As such, these results suggest that rLFA is moderating the relationship between affective arousal and behavioral response rather than the relationship between emotional stimulus and affective arousal. This finding supports the efferent model.

Papousek et al. (2014) also found that those who had a higher decrease in rLFA in response to videos of injury and death had greater mood deterioration and film-related intrusive memories and avoidance over the next week. This study suggests that greater change in rLFA may be a vulnerability factor for negative affect and related symptoms. Given the research discussed earlier that suggests changes in rLFA related to emotional stimuli are adaptive (Papousek & Schulter, 2004), future research may examine under what conditions changes in rLFA may be adaptive or maladaptive.

Despite the implication of rLFA as a potential moderator of emotional response, results are somewhat inconsistent and should be further explored. Hagemann, Naumann, Becker, Maier, and Bartussek (1998) were able to replicate the results of Tomarken et al. (1990) with a Cz reference but not with a linked mastoid reference. Coan and Allen (2003b) explicitly tested rLFA as a moderator of emotional response by asking participants to make disgust, anger, joy, sadness, and fear faces. They found that rLFA at rest moderated the relationship between directed faces and self-reported emotional response; such that more rLFA predicted increased experience of anger, joy, and disgust. More rLFA has been associated with anger and joy in the literature more broadly (Carver & Harmon-Jones, 2009; Harmon-Jones & Allen, 1998; Killeen & Teti, 2012), whereas disgust has been associated with less rLFA (Jones & Fox, 1992).

Some evidence suggests that, in line with the capability model, rLFA may be a more consistent moderator of emotional response when measured during emotional stress task rather than at rest. For example, Goodman, Rietschel, Lo, Costanzo, and Hatfield (2013) studied the relationship between eyeblink startle magnitude and threat of shock with both rLFA at rest and during an emotional challenge task. rLFA at rest did not serve as a moderator of this relationship, but rLFA during the threat of shock did serve as a moderator of this relationship. Results suggest that rLFA measured during stress situations may predict emotion regulation abilities. Additionally, these results suggest that the relationship between rLFA and emotion regulation may not be as robust at rest as during emotional challenge tasks, as suggested by the capability model.

In addition to emotional response, a few studies have also examined rLFA as a moderator of attention processes. Amodio (2010) explored rLFA as a predictor of attention to black versus white faces. The researchers measured rLFA during a goal-directed state in line with the capability model. They found that those with more rLFA during these periods had larger P2 responses to black versus white faces, suggesting rLFA as a moderator of attention response to race. Interestingly, they also conducted a more in-depth moderated mediation model by examining the relationship between P2 and action control. Whereas frontal asymmetry predicts P2 amplitude for all participants, P2 amplitude predicts action control only for low-prejudice participants. In this latter analysis, frontal asymmetry is the predictor, P2 amplitude is the third variable, and action control is the outcome.

More recently, Tullett, Harmon-Jones, and Inzlicht (2012) found that individuals with less rLFA responded with more empathic concern to images related to a charity than those with more rLFA. This finding suggests that rLFA may not only be a moderator of emotional response but also empathic responding.

Hall, Ekkekakis, and Petruzzello (2010) also examined emotional responses to exercise using rLFA. They found that less rLFA at rest predicted lower energetic arousal at multiple points after exercise. This finding suggests that lower rLFA may predict less energetic arousal in response to approach-related tasks such as exercise.

A number of studies have also investigated the relationship between depressive symptoms and other trait variables, with rLFA as a moderator. For example, Kline and Allen (2008) examined the relationship between defensiveness and depressive symptoms. They found that rLFA and defensiveness interacted to predict depressive symptoms. Those individuals with less rLFA had a positive relationship between defensiveness and depressive symptoms, whereas those individuals with more rLFA had a negative relationship between them. As such, rLFA may in this case serve as an antagonistic moderator, as the relationship between defensiveness and depressive symptoms is reversed as a function of frontal asymmetry. Lopez-Duran, Nusslock, George, and Kovacs (2012) examined the relationship between stressful life events and internalizing symptoms in children at risk for depression based on their mother's depression. They found that rLFA serves as a moderator of the relationship between stressful life events and internalizing symptoms. Specifically, the less rLFA one has, the more an increase in stressful life events leads to an increase in internalizing symptoms.

Given the relationship between rLFA and psychopathology discussed above (Allen & Reznik, 2015; Mathersul et al., 2008), a few studies have examined rLFA as a potential moderator of treatment response. Exploration of rLFA as a moderator of treatment response would theoretically suggest that treatments work only for those who have certain levels of rLFA. Bruder et al. (2001) found that rLFA moderated response to fluoxetine in individuals with depressions. Those participants with less rLFA did not respond to treatment, whereas those with more rLFA responded to treatment. However, Gollan et al., (2014) examined rLFA as a potential moderator of behavioral activation treatment. As behavioral activation aims to increase approach-related behavior, it is likely that it would impact systems of approach/withdrawal motivation (Davidson, 2000). Yet, no relationship between rLFA and treatment response was found. As this study examined rLFA at rest, future research may examine rLFA as a moderator of treatment response during an emotionally evocative task. It also may be possible that rLFA will only serve as a moderator of treatment response for certain populations.

In a multicenter, randomized trial, Arns et al. (2015) examined 2 min of baseline resting EEG as a predictor of treatment response to escitalopram, sertraline, or venlafaxine-extended release. They found that more rLFA in women predicted response to escitalopram and sertraline but not to venlafaxine. However, the authors failed to find a relationship between rLFA and depression status (1,008 depressed participants compared to 336 healthy controls) possibly due to the short length of baseline recording (2 min) and using the average reference. Future research may examine whether sex or gender plays a role in identifying treatment moderators.

In another study, Moscovitch et al. (2011) examined rLFA at 8–10 Hz, 10–13 Hz, and the typical 8–13 Hz as potential predictors of treatment response for individuals with social anxiety disorder receiving cognitive behavioral therapy. In this study, rLFA moderated treatment response, such that those with more rLFA at rest at 10–13 Hz before treatment had greater reduction in social anxiety from pre- to posttreatment. Although these results are promising, further research may examine potentially differential effects using low, high, and full alpha bands as well as traditional rLFA versus other EEG profiles as moderators of treatment response.

Additionally, some research that has addressed alpha activity as a moderator of treatment response has tended to utilize alpha asymmetry at nonfrontal sites, or using other frequency bands. These studies are mentioned here only because they are often cited as studies examining frontal alpha asymmetry specifically. For example, Bruder et al. (2008) found that measures of alpha asymmetry at occipital, rather than frontal, sites predicted response to selective serotonin reuptake inhibitors (SSRIs) in individuals with depression. Another study found that baseline theta power and relative theta power during the first week predicted treatment response to SSRIs or venlafaxine but frontal asymmetry did not (Iosifescu et al., 2009). It was also found that change in combined theta and alpha power after SSRI treatment significantly predicted change in suicidal ideation from baseline (Iosifescu et al., 2008). Additionally, Quraan et al. (2014) examined alpha activity as a predictor of treatment response for individuals with depression undergoing deep-brain stimulation surgery. Both theta at frontal sites and alpha activity at parietal sites predicted treatment response, revealing that theta and alpha activity may serve as moderators in predicting treatment response.

1.6 Frontal asymmetry as a mediator of emotion

A mediator is a third variable through which (or partially through which) the predictor influences the outcome variable (Baron & Kenny, 1986). Whereas moderators allow insight into when or for whom a relationship holds, mediators provide insight into how or why a process occurs. For example, rLFA would serve as a mediator in a case where a given treatment alters rLFA, which in turn alters depressive symptoms. For example, Allen, Harmon-Jones, and Cavender (2001) used biofeedback to alter rLFA, which in turn altered affect. Frontal asymmetry serves as a mediator in this case because it is related to both biofeedback and outcome; it is able to explain variance in the relationship between the predictor and outcome. Mediator analyses may occur at any level (psychological, social, neural, etc.) and can improve our theoretical understanding at any of them. A mediator may or may not serve as a direct index of the mechanism by which actual change occurs. As such, it may be important to consider the distinction between mediators and mechanisms of change (Kazdin, 2007). In examining rLFA as a potential mediator, it is important to note that rLFA may be conceptualized as indexing either a psychological construct or the activity of specific neural systems or both. As a mediator, rLFA may provide a theoretical frame to examine how the predictor impacts the outcome, but should not be held inherently as the mechanism by which change occurs. Figure 2 illustrates the conceptual and statistical relationships in the case of mediation.

According to Baron and Kenny (1986), mediation occurs when (a) variations in the predictor account for variations in the third variable, (b) variations in the third variable account for variations in the outcome, and (c) when the relationship between predictor and third variable as well as third variable and outcome are adjusted for, the relationship between predictor and outcome should disappear. In other words, frontal asymmetry would serve as a mediator when some change in frontal asymmetry is necessary for the predictor to affect the outcome. One relationship may include multiple mediators. Baron and Kenny (1986) propose using multiple regression models to identify mediators. Although this technique has been widely used, Preacher and Hayes (2004) suggest use of the Sobel test. The Sobel test is a t test that compares the null hypothesis that mediator has no effect on the effect of the mediator. More recently, Preacher and Hayes have developed a computation to conduct mediation using path analysis called PROCESS (Hayes, 2012, 2013). Researchers may also explore specialized techniques for longitudinal data, such as panel models and latent growth curve models (Preacher, 2015).

While mediators and confounds share the same statistical relationships, they can be distinguished using theory to guide analysis. Mediators are theorized to be related to the cause by which a process likely occurs and are of significant theoretical interest or importance. On the other hand, confounders are either variables that are related by chance and cannot reasonably be the cause by which the predictor affects the outcome or fail to have significant theoretical interest or importance a priori. For example, imagine a novel psychosocial intervention proposed to impact rLFA. The researchers hypothesize that increasing rLFA will decrease depressive symptoms. In the case where the intervention (vs. control condition) increases rLFA and that change in rLFA accounts for decrease in depressive symptoms, rLFA would serve as a mediator. The findings would support the researchers' hypothesis that rLFA may be changed though this intervention to decrease depressive symptoms. However, one can imagine the case where the intervention does not alter rLFA, yet depressive symptoms are decreased. The perplexed researchers may then search for unexpected mechanisms that caused this change in depressive symptoms. Because the researchers collected a host of other potentially confounding variables, they compare the impact of intervention versus control on all of these variables. They find that the intervention significantly increased participants' time out of the house compared to the control condition. The time out of the house variable also accounted for decrease in depressive symptoms. In this latter case, the proposed mediator (rLFA) does not exist, and the researchers have identified a confounder (amount of time out of the house). Many researchers may predict that increasing time out of the house would decrease depressive symptoms, and this discovery provides important novel information for the researchers; yet, in this particular study, this variable would serve as a confound due to its lack of significant theoretical relevance to the intervention's impact on rLFA. In this way, researchers can distinguish cases where rLFA is spuriously related and causally related to the outcome and predictor (MacKinnon et al., 2013).

1.7 Evidence supporting rLFA as mediator of emotion

Studies designed to examine rLFA as a statistical mediator can lead to the identification of mechanisms by which emotional manipulations, interventions, or individual differences cause changes in mood and emotion. Because rLFA can be both an index of psychological and neural mechanisms, examining rLFA as a mediator offers substantial potential to broaden theories of emotion regulation and psychopathology by identifying potential psychological, social, or neural mechanisms of change. See Table 4 for studies examining rLFA as a mediator of emotion and psychopathology.

In order to establish rLFA as a mediator, it is necessary to determine that the rLFA score is both related to the predictor variable and the outcome variable. The extent to which the predictor variable alters rLFA should correspond to the extent of change in the outcome variable. For example, imagine that a researcher aims to examine rLFA as a mediator of depression treatment response. In the case that rLFA serves as a true mediator, rLFA scores should change based on the amount of intervention, and decrease in depressive symptoms should also be related to change in rLFA. As such, although a number of studies have elicited changes in rLFA (Coan & Allen, 2003b; Harmon-Jones, Gable, & Price, 2011), few studies have examined the impact of such changes on subsequent outcome measures like self-reported mood or behavior (Allen et al., 2001). Further research examining outcomes remains needed. In particular, researchers may examine whether changes in rLFA correspond to changes in clinical self-report.

In studies without measured outcomes, frontal asymmetry must be conceptually considered an outcome rather than a mediator variable. For example, Barnhofer et al. (2007), discussed in detail above, found that rLFA changed in response to treatment type for previously suicidal individuals but did not link these changes to any psychological phenomena. It is possible that rLFA serves as a mediator in this case between intervention and depressive symptoms or positive affect. As there is no outcome variable explicitly measured in this study, rLFA functions as the outcome variable. Additionally, one study examined rLFA as a potential mediator of depression using cortisol (Tops et al., 2005). It was found that administering cortisol to 11 healthy male volunteers decreased rLFA. Because less rLFA is associated with depression and increased cortisol is associated with depression, this research may suggest that rLFA is a mediating pathway in the onset of depression. However, this study did not conduct a true mediation analysis, as an outcome variable that would likely be related to the change in rLFA was not examined. Evidence to date supports the general stability of rLFA across episodes of major depressive disorder (MDD), both within subjects over time for those with history of MDD (Allen, Urry, Hitt, & Coan, 2004; Vuga et al., 2006) and between subjects differing in clinical status (Stewart et al., 2010). Yet, identifying that a treatment impacted depression via rLFA (as a mediator of depression treatment response) would have significant implications. Such treatments may be most promising for providing lasting remission and reducing risk for future episodes.

Additionally, Allen, Mcknight, Moreno, Demaree, and Delgado (2009) examined rLFA in response to tryptophan depletion. They found that change in rLFA because of tryptophan depletion predicted the development of depression 6 to 12 months out. Although there was no main effect of tryptophan depletion on depression status, this study suggests a potential causal role of change in rLFA to depression.

At the time of publication of Coan and Allen (2004), they reported that there had been no full meditational analyses (i.e., each path tested) involving rLFA. That paper summarized the results of a then recent biofeedback study that altered frontal asymmetry (Allen et al., 2001), in which biofeedback altered rLFA as well as self-reported affect and facial muscle activity in response to emotionally evocative films. There was no formal mediation analysis, however, and rLFA may be implicated as a potential mediator if the extent to which rLFA changed corresponds to the extent to which self-reported mood changed. They found that mood and electromyography (EMG) effects were stronger in those whose EEG asymmetry scores changed the most in response to the biofeedback training providing evidence for rLFA mediation. Since that time, Peeters, Ronner, Bodar, van Os, and Lousberg (2014) used a single session of neurofeedback to alter frontal asymmetry scores in 49 women and conducted a mediation analysis. Although rLFA was altered, there were no self-reported changes in mood. Therefore, it is unclear under what conditions changes in rLFA may influence mood changes. Additionally, although neurofeedback may alter rLFA, it is unclear from this research what impacts it may have on mood or motivation. An important distinction may be whether unprovoked mood (as in Peeters et al., 2014) or whether emotional reactions to evocative stimuli is assessed (as in Allen et al., 2001). Further research may attempt to provide additional theoretical clarity as to the conditions under which altering rLFA changes mood or other clinical outcomes.

A number of studies have elicited changes in rLFA and examined the effects on a meaningful outcome variable. For example, Peterson, Shackman, and Harmon-Jones (2008) found that hand contraction elicited more rLFA and increased behavioral aggression on a subsequent task. In this study, rLFA serves a mediator of the relationship between hand contraction and aggression. This study supports the finding that increases in rLFA may play a causal role in aggression. A more recent follow-up study also found that rLFA was a partial mediator of the relationship between hand contractions and anger induced by an ostracism task (Peterson, Gravens, & Harmon-Jones, 2011). Another study examined rLFA as a mediator of the relationship between stress and aggression (Verona, Sadeh, & Curtin, 2009). It found that stress induction in the laboratory increased rLFA. Frontal asymmetry scores after the stress induction also predicted aggression in the laboratory, using a task in which participants are asked to shock an employee (no shocks are actually delivered).

Amodio, Devine, and Harmon-Jones (2007) also explored the relationship between guilt and rLFA. After a multiracial facial viewing task, they provided participants with false performance feedback that they had made anti-Black responses. They found that to the extent that guilt increased, rLFA decreased.

A number of recent studies have also expanded theory by examining rLFA as a longitudinal predictor. Lewis, Weekes, and Wang (2007) studied the relationships between rLFA, psychological stress, hormonal stress, and negative health during both high and low stress times for undergraduates. They found that rLFA shifted from more rLFA during periods of low academic stress to less rLFA during periods of high academic stress. Importantly, greater change in rLFA was related to increases in negative health from the low to high stress periods. These results suggest that rLFA may mediate the relationship between environmental stress and negative health outcomes; the more an individual's rLFA is altered by periods of stress, the more negative health outcomes one is likely to experience. This study is innovative in its ability to link rLFA as a mediator of the relationship between emotional and physical health in a natural environment for the first time.

1.8 The importance of moderated mediation and mediated moderation

1.8.1 Moderated mediation

In addition to mediation and moderation effects, more complex models of mediated moderation and moderated mediation may be relevant to frontal asymmetry. Moderated mediation occurs when the mediation process of a given relationship differs as a function of a moderator (e.g., group or condition). The moderator may influence the relationship between the predictor and the mediator or the relationship between the mediator and the outcome. For example, imagine that moderated mediation exists such that gender serves as a moderator for a cognitive behavioral therapy (CBT) that reduces cognitive biases and, therefore, depressive symptoms. Gender may influence the link between CBT and cognitive biases, such that women have further reduction in cognitive biases after CBT than men, or gender may influence the link between cognitive biases and depressive symptoms, such that women have further reduction in depressive symptoms due to reduction in cognitive biases. There does not need to be an overall moderation effect for a mediated moderation effect to occur (Muller, Judd, & Yzerbyt, 2005). For example, imagine that a novel treatment improves depression scores. The novel treatment includes behavioral activation, mindfulness, and cognitive bias modification. For those with less rLFA, behavioral activation serves as the mediator. For those with more rLFA, mindfulness serves as the mediator. As such, the mediating process differs as a function of one's rLFA score (the moderator). Figure 3 shows an example of such a moderated mediation, which represents just one of many ways that a moderator may impact mediation. Importantly, there should be theoretical backing for such a model (e.g., depression in those with less rLFA is primarily an approach-related deficiency that may be improved with behavioral activation) but different mediating processes for different groups may emerge unexpectedly. Amodio et al. (2007) recently found moderated mediation in the relationship between frontal asymmetry and action control. P2 was found to be a mediator between rLFA and action control but only for participants with positive racial attitudes.

1.8.2 Mediated moderation

Mediated moderation occurs when there is a mediated process through which moderation occurs (Muller et al., 2005). In contrast to moderated mediation, mediated moderation occurs only when moderation occurs. For example, imagine again that behavioral activation improves depression scores only for individuals with lower BAS scores. A moderated mediation would occur if behavioral activation therapy only reduces depressive symptoms in individuals with low BAS scores because those with low BAS scores have less rLFA. As those with low BAS scores have less rLFA, rLFA mediates the relationship between treatment and outcome but also is moderated by BAS score (see Figure 4 for conceptual representation of mediated moderation). Importantly, recent research has combined mediated moderation and moderated mediation under the term conditional process analysis, and such models can be evaluated using PROCESS software (Hayes, 2012). Hayes (2013) argues that the semantic distinction between moderated mediation and mediated moderation does not provide conceptual benefit to modern mediation and moderation research. In particular, he argues that reframing interpretation in terms of “moderated mediation” will lend to clearer results than “mediated moderation.” Although the ability and knowledge to distinguish these two types of models may always be relevant to interpreting prior research, we argue that clear, interpretable explication of models with reference to variables that moderate and mediate, when done with great care and effort, may supersede the terms moderated mediation and mediated moderation without explication. Hayes (2016) may provide the reader with a more thorough understanding of conditional process models and ability to run and clearly explicate such models.