A Preliminary Investigation of ERP Components of Attentional Bias in Anxious Adults using Temporospatial Principal Component Analysis

Abstract

Threat-related attention bias is thought to contribute to the development and maintenance of anxiety disorders. Dot-probe studies using event-related potentials (ERPs) have indicated that several early ERP components are modulated by threatening and emotional stimuli in anxious populations, suggesting enhanced allocation of attention to threat and emotion at earlier stages of processing. However, ERP components selected for examination and analysis in these studies vary widely and remain inconsistent. The present study used temporospatial principal component analysis (PCA) to systematically identify ERP components elicited to face pair cues and probes in a dot-probe task in anxious adults. Cue-locked components sensitive to emotion included an early occipital C1 component enhanced for happy versus angry face pair cues and an early parieto-occipital P1 component enhanced for happy versus angry face pair cues. Probe-locked components sensitive to congruency included a parieto-occipital P2 component enhanced for incongruent probes (probes replacing neutral faces) versus congruent probes (probes replacing emotional faces). Split-half correlations indicated that the mean value around the PCA-derived peaks were reliably measured in the ERP waveforms. These results highlight promising neurophysiological markers for attentional bias research that can be extended to designs comparing anxious and healthy comparison groups. Results from a secondary exploratory PCA analysis investigating the effects of emotional face position and analyses on behavioral reaction time data are also presented.

1. Introduction

Threat-related attentional bias, defined as the preferential tendency to allocate attention toward or away from threatening stimuli (Mogg & Bradley, 2018), developed as a consequence of humans’ evolution in environments where dangers constantly threatened survival and reproductive advantage (Ohman, Flykt, & Esteves, 2001). Cognitive and neural research has also established that attentional processing of threatening stimuli, such as angry facial expressions, are prioritized over neutral stimuli in healthy individuals (Maratos, Mogg, & Bradley, 2008; Okon-Singer, 2018; Yiend, 2010). However, threat-related biases can also become maladaptive and contribute to the development and maintenance of anxiety disorders, which are associated with hypervigilance to potential threat in preparation for future danger, cautious or avoidant behaviors (Cisler & Koster, 2010; DSM-5 American Psychiatric Association, 2013; Mogg, Bradley, Williams, & Mathews, 1993), and delayed disengagement from threat (Amir, Elias, Klumpp, & Przeworski, 2003).

Behavioral attentional bias research has demonstrated that threat-related biases are often observed in anxious populations (Bar-Haim, Lamy, Pergamin, Bakermans-Kranenburg, & van Ijzendoorn, 2007; Cisler & Koster, 2010; Mogg & Bradley, 1998). However, behavioral measures (e.g., reaction times (RTs)) lack reliability and sensitivity, provide an indirect measure of attentional processing (Horley, Williams, Gonsalvez, & Gordon, 2004), and can be confounded by post-perceptual processes such as motor responses and decision making (Handy, Green, Klein, & Mangun, 2001; Mueller et al., 2009). Additionally, behavioral measures are not-well suited to examining the components, or observable and measurable characteristics, of threat-related attentional bias arising at different stages of information processing: (1) facilitated attention to threat, or hypervigilance (i.e., the relative ease or speed with which attention is initially and involuntarily drawn to a threat stimulus), during early, automatic processing stages, (2) difficulty disengaging attention away from threat (i.e., the degree to which a threatening stimulus captures attention and impairs switching attention from the threatening stimulus to another stimulus), and (3) attentional avoidance of threat (i.e., automatic or strategic shifting of attention away from the spatial location of threat, even when the threatening item is no longer present) during early or late processing stages (Cisler & Koster, 2010; Gupta, Kujawa, & Vago, 2019). However, neurophysiological measures, such as event-related potentials (ERPs), are well-suited for investigating threat-related attentional bias and its components because these measures allow for the examination of the time course of attention to threat with millisecond resolution (Kappenman, Farrens, Luck, & Proudfit, 2014; Kappenman, MacNamara, & Proudfit, 2015).

Several ERP components, including the early C1, P1, N1, N170, P2, N2, and N2pc and the later P3 and late positive potential (LPP), have been analyzed in the attentional bias literature (Gupta et al., 2019). The C1, peaking 80–100 milliseconds (ms) poststimulus at posterior midline sites (Luck, 2014), is triggered by the appearance of a stimulus in the visual field (Clark & Hillyard, 1996; Eldar, Yankelevitch, Lamy, & Bar-Haim, 2010; Luck, Woodman, & Vogel, 2000). The P1, peaking 100–130 ms poststimulus at lateral occipital sites (Luck, 2014), is sensitive to allocation of attention to stimuli (Clark & Hillyard, 1996). The N1, consisting of several subcomponents peaking between 100–200 ms poststimulus at anterior and posterior sites, is influenced by spatial attention and discrimination of attended stimuli (Luck, 2014). The N170, peaking approximately 170 ms at lateral occipital sites (Luck, 2014), is regarded as a face-specific ERP component (Bentin, Allison, Puce, Perez, & McCarthy, 1996). The P2 reflects allocation of attentional resources during the processing of emotional facial expressions (Bar-Haim, Lamy, & Glickman, 2005; Eldar et al., 2010; Torrence & Troup, 2018) but can be difficult to distinguish from the N1, N2, and P3 at posterior sites (Luck, 2014). The N2 consists of several subcomponents; the posterior N2 subcomponent has been associated with discrimination and classification of visual stimuli (Luck, 2014). The N2pc, occurring 200–300 ms poststimulus at posterior scalp sites contralateral to an attended object, measures whether attention is covertly directed to a stimulus (Luck, 2014). The P3 component, peaking 350–600 ms poststimulus (Luck, 2014), includes the frontal P3a component and a parietal P3b component; both are elicited by unpredictable, infrequent stimulus changes. The LPP, a central-parietal, midline component occurring approximately 300 ms poststimulus, is larger following the presentation of emotional compared to neutral stimuli (Hajcak, Dunning, & Foti, 2009) and can be sustained for several seconds after emotional stimuli are presented (Hajcak et al., 2009).

A review on the neural chronometry of threat-related attentional bias showed that early ERP components, including the P1, N170, P2, and N2pc, are modulated by emotional (e.g., threatening or positive) stimuli in both healthy and anxious populations, suggesting that both groups display enhanced allocation of attention to emotional stimuli at earlier stages of processing. However, later components (e.g., P3) are modulated by emotional stimuli more reliably in healthy, compared to anxious, populations. Thus, healthy populations show more evidence for conscious, evaluative processing of threat and emotion and disengagement difficulties at later stages of processing (Gupta et al., 2019).

The dot-probe task (MacLeod, Mathews, & Tata, 1986) is frequently used in both behavioral and ERP studies to assess attentional bias in spatial orienting to threatening cues (Mogg & Bradley, 2016). In the task, two visual stimuli (e.g., words, faces, scenes), called cues, are briefly and simultaneously presented above and below or to the left and right of a fixation cross. One cue is emotional or threatening and the other is neutral. After the cues disappear, a probe, or target (e.g., a dot or bar), appears in the spatial location of one of the cues. Participants must quickly and accurately respond to the location or identity of the probe. Faster RTs to probes are observed when they occur in the attended rather than the unattended location (Navon & Margalit, 1983). Thus, participants displaying attentional bias toward threat will typically demonstrate faster RTs to probes appearing in the location of threatening, compared to neutral, stimuli (Bar-Haim et al., 2007; Van Bockstaele et al., 2014).

Although the dot-probe task is widely used, there is evidence that behavioral measures of attentional bias are not reliably elicited; thus, there has been a growing interest in improving the metrics obtained from the task through the inclusion of ERPs (Kappenman et al., 2014; Torrence & Troup, 2018). In ERP studies, components time-locked to the presentation of cues and probes are analyzed. Amplitude or latency modulations of cue-locked ERPs may reflect attentional bias occurring at early stages of processing, whereas modulations of probe-locked ERPs may reflect attentional bias occurring at later stages of processing (Gupta et al., 2019).

Dot-probe ERP studies with anxious populations have examined a wide range of components yielding a variety of results (see Gupta et al., 2019). Several of these studies have focused on early components such as the P1. For example, Helfinstein, White, Bar-Haim, & Fox (2008) observed that individuals with high social anxiety displayed higher mean P1 amplitudes to angry-neutral face pairs compared to individuals with low social anxiety, suggesting increased sensory processing of faces in individuals with high levels of social anxiety. Similarly, Mueller et al. (2009) demonstrated that, compared to healthy controls, participants with social anxiety disorder (SAD) displayed enhanced P1 amplitudes to angry-neutral versus happy-neutral face pairs and decreased P1 amplitudes to probes replacing emotional (angry and happy) versus neutral faces, suggesting an early hypervigilance to angry faces and reduced visual processing of emotionally salient locations at later stages of information processing in SAD participants, respectively. However, other early components have also been examined. For example, Eldar et al. (2010) demonstrated that, compared to non-anxious individuals, anxious individuals displayed enhanced occipital P2 amplitudes in response to face displays, regardless of whether the facial emotion was angry, happy, or neutral. This P2 modulation in anxious individuals serves as an indicator of attentional commitment to processing facial emotional expressions. Rossignol, Campanella, Bissot, & Philippot (2013) observed that individuals with high social anxiety displayed enhanced P2 amplitudes in response to angry-neutral compared to fear-neutral face pairs, suggesting enhanced allocation of attention to angry faces. Additionally, Fox, Derakshan, & Shoker (2008) studied high trait anxiety and low trait anxiety groups and observed that angry expressions elicited an enhanced N2pc, but only in participants reporting high levels of trait anxiety, suggesting that participants with high trait anxiety exhibit rapid exogenous orienting of spatial attention to threatening cues.

Although neural measures, such as ERPs, are thought to be more reliable than behavioral measures, such as RTs, the scoring of ERP components varies considerably across studies. For example, while Helfinstein et al. (2008) examined the P1 mean amplitude averaged across electrodes O1 and O2 95–140 ms after face onset, Mueller et al. (2009) measured the P1 as the most positive peak in the time window of 80–150 ms following face or probe onset at electrodes PO7 and PO8. Similarly, Eldar et al. (2010) quantified the P2 as the mean amplitude over electrodes O1 and O2 195–250 ms after face display onset, while Rossignol et al. (2013) measured the P2 as the mean amplitude 240–400 ms after face pair presentation at electrodes O1 and O2. Further, to our knowledge, prior work has yet to characterize the reliability of early emerging ERPs in dot-probe tasks.

The aforementioned results clearly demonstrate that a variety of cue- and probe-locked ERP components have been examined in dot-probe studies; however, the choice of which components to focus on to capture attentional bias is inconsistent across the literature. Additionally, the scoring and labeling of ERP components in attentional bias studies varies widely. Thus, a special method is required to more systematically select ERP components sensitive to early and later stages of processing in dot-probe tasks. Principal component analysis (PCA) provides an effective way to analyze high-density ERP datasets and to separate components that vary in their sensitivity to spatial, temporal, or functional parameters (Dien & Frishkoff, 2005). PCA has successfully been utilized to differentiate ERPs sensitive to emotion in a variety of studies using other paradigms besides the dot-probe (Kujawa, Weinberg, Hajcak, & Klein, 2013; Mulligan, Infantolino, Klein, & Hajcak, 2020; Pegg et al., 2019).

The present preliminary study used temporospatial PCA to systematically identify the timing and scalp distributions of ERPs elicited to angry-neutral and happy-neutral face pair cues and bar probes in a dot-probe task adapted from Mueller et al. (2009) in adults with moderate to high levels of anxiety. The present analyses are part of a larger, ongoing study investigating the neurobiological mechanisms of mindfulness-based cognitive therapy (MBCT) [NCT03571386], including effects on ERP markers of threat-related attentional bias in anxious populations. The PCA analyses were performed on the dot-probe task administered to a sample population with moderate to high levels of anxiety prior to an MBCT intervention. The PCA results will not only inform which cue- and probe-locked components to examine in further analyses associated with the ongoing clinical intervention study, but also clarify scoring windows and electrode sites to use for each of these components. Subsequently, we sought to determine whether the mean value around the PCA-derived peaks were reliably measured in our ERP waveforms using the Spearman-Brown split-half reliability method. A secondary exploratory PCA analysis was conducted to determine the extent to which ERPs derived through PCA were moderated by the position of emotional (angry or happy) faces in the face pair cues. Accounting for the location of the emotional face in this analysis allowed us to identify whether the N2pc was elicited to emotional faces, as this ERP arises at posterior scalp sites contralateral to an attended object (Luck, 2014). Behavioral analyses were also conducted to determine the effects of emotion and congruency on reaction times in the dot-probe task. While the present study lacks a healthy comparison group, identification of components in the anxious sample using temporospatial PCA will yield promising and reliable ERP measures of attentional bias that can be extended to designs comparing anxious and healthy comparison groups.

2. Methods

2.1. Participants

In order to participate in the larger treatment study, individuals had to (1) be between the ages of 18 and 55 years, (2) have moderate to high levels of anxiety, indexed by a score of 40 or above on the State-Trait Anxiety Inventory, Trait Scale (STAI-T) (Spielberger, Gorsuch, Lushene, Vagg, & Jacobs, 1983), and (3) be considered stable on maintenance anti-anxiety, anti-depression, or as needed (PRN) medications for at least one month prior to enrollment. Individuals were excluded if they had (1) Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) diagnostic criteria indicating bipolar I or II, dementia, psychotic, borderline, or narcissistic personality disorders, (2) a current history (≤ 6 months) of regular meditation practice (> 1 session per week; > 10 minutes per session), (3) a current history (≤ 6 months) of substance abuse and/or dependence, (4) an inability to understand English at a level necessary for informed consent and understanding instructions, or (5) a serious underlying systemic or comorbid disease precluding physical or cognitive ability to participate. Individuals were asked not to start individual psychotherapy or a regular meditation or yoga practice during the treatment study.

In the present study, 25 anxious adults (21 females, 4 males) with a mean age of 32.12 years (SD = 10.05) were recruited from the greater Nashville community using ResearchMatch and the Vanderbilt University Medical Center research notification distribution listserv. Participant racial breakdown included 4% Asian, 8% Black or African American, 84% White, and 4% were more than one race. Participants’ ethnicity included 12% Hispanic or Latino. Participants provided written informed consent and received monetary compensation for their participation. The study was approved by the Vanderbilt University Institutional Review Board.

2.2. Dot-Probe Task

A dot-probe task adapted from Mueller et al. (2009) (see Figure 1) with simultaneous EEG recording was used to assess threat-related attentional bias in the anxious participants.

Figure 1.

Schematic of the dot-probe task used in the present study.

2.2.1. Stimuli

Pairs of face stimuli were created using grayscale photographs of males and females portraying angry, happy, and neutral facial expressions from Ekman’s Pictures of Facial Affect (Ekman & Friesen, 1976). All of the happy face stimuli used in the present study exhibited smiles with exposed teeth, while half of the angry faces used in this study featured exposed teeth and the other half featured compressed lips. Each face pair consisted of two different identities of the same sex portraying a neutral expression and either an angry or happy facial expression. This yielded four conditions: angry-neutral, neutral-angry, happy-neutral, and neutral-happy. Each emotional expression appeared equally often to the left or right of the neutral expression. Faces were cropped into 8 centimeter (cm) × 10 cm ovals and set on a black background. The centers of the faces were 18 cm apart. The faces were presented in the upper visual field and were viewed at a distance of 70 cm. The probe was a white, vertical rectangular bar measuring 6 cm × 0.4 cm and was presented on either the left or right side of the screen in the same upper visual field location as the faces. The fixation cross measured 2 cm × 2 cm with a thickness of 0.1 cm and was presented centrally on the lower part of the screen. All stimuli were set on a black background and presented on a 24-inch monitor with a Dell desktop computer running E-Prime (Psychology Software Tools, Pittsburgh, PA). Participants made responses to the stimuli using a Cedrus RB-844 button box (Cedrus, San Pedro, CA).

2.2.2. Procedure

The dot-probe task began with a practice block of 16 trials followed by six blocks of 120 trials each (720 trials total). Each block was separated by a short rest break. Each trial began with the presentation of a fixation cross for 250 ms followed by presentation of the face pair cues for 100 ms. The interstimulus interval varied randomly from 200 to 300 ms (in 25 ms increments); thus, the stimulus onset asynchrony was 300–400 ms. The probe then appeared for 150 ms in either location previously occupied by a face. The intertrial interval was 1250 ms. Female face pairs were presented 60% of the time, and male face pairs were presented 40% of the time. Happy and angry face pairs appeared equally often and with equal frequency in the right and left visual field. Probes also appeared with equal frequency in the right and left visual field. All stimuli were randomized and counterbalanced across participants. For each trial, participants were instructed to focus on the fixation cross while concurrently monitoring the location of the probe. Participants were asked to press one of two buttons on the response box to indicate which side of the screen the probe was on. Response times were recorded from probe onset. Accuracy was measured as the number of correct responses (“hits”) and the number of incorrect responses (“misses”). Trials with incorrect responses, response times <100 ms, or response times >1500 ms were excluded from behavioral analyses.

2.3. Power Analysis

In order to determine whether we were sufficiently powered to detect emotion and congruency effects, we performed a post-hoc power analysis in G*Power 3.1 (Faul, Erdfelder, Lang, & Buchner, 2007) using a sample size of 25, alpha of .05, and within-subjects emotion and congruency effect sizes from Mueller et al. (2009), as our experimental design was based closely on this study. Results indicated that we were well-powered to detect within-subjects emotion and congruency effects (power estimate > 0.99).

2.4. EEG Recording and Data Reduction

EEG was recorded continuously using Brain Vision Recorder (Brain Products GmbH, Gilching, Germany), BrainAmp DC (Brain Products GmbH, Gilching, Germany), and a 64-channel actiCAP (Brain Products GmbH, Gilching, Germany) with a sampling rate of 500 Hz and an FCz reference. Electrodes Fp1, Fp2, FT9, and FT10 were removed from the cap and used as EOG channels; vertical eye movements were recorded using electrodes placed above and below the left eye, and horizontal eye movements were recorded using electrodes placed near the outer canthus of each eye. Impedance of all channels was kept below 10 kΩ.

Data were processed using Brain Vision Analyzer (Brain Products GmbH, Germany). Data were first filtered between 0.1–30 Hz via zero-phase shift band-pass (IIR Butterworth) and 60 Hz notch filters and were subsequently re-referenced offline to an average reference, yielding 61-channel EEG data (the original reference channel, FCz, was regained as a data channel). Raw data inspection was performed on the continuous EEG data to identify and mark artifacts. Ocular artifacts were corrected using the regression method (Gratton, Coles, & Donchin, 1983). When required, topographic interpolation by spherical splines was performed.

For the cue condition, data were segmented into (1) trials where angry-neutral face pairs were presented, and (2) trials where happy-neutral face pairs were presented. For the probe condition, data were segmented into (1) presentation of angry congruent probes (i.e., probe replaces angry face in angry-neutral face pairs), (2) presentation of angry incongruent probes (i.e., probe replaces neutral face in angry-neutral face pairs), (3) presentation of happy congruent probes (i.e., probe replaces happy face in happy-neutral face pairs), and (4) presentation of happy incongruent probes (i.e., probe replaces neutral face in happy-neutral face pairs). All segments were extracted beginning 50 ms before and ending 300 ms after stimulus presentation. Cue- and probe-locked segments were baseline corrected from −50 to 0 ms. Artifact rejection was completed using semi-automatic inspection, individual channel mode, and the following criteria: maximal allowed voltage step: 50 μV/ms; maximal allowed difference of values in intervals: 200 μV (interval length: 200 ms); and lowest allowed activity in intervals: 0.1 μV (interval length: 100 ms). Artifact rejection also removed trials where voltages exceeded +/− 75 μV. Only trials where participants responded correctly were used to calculate each subject’s averages, and subsequently, the grand averages. Subject averages and grand averages were computed with individual channel mode enabled. The mean number of trials included in the grand averages at electrode PO4 were as follows: angry cue: 349 trials; happy cue: 351 trials; angry congruent probe: 173 trials; angry incongruent probe: 176 trials; happy congruent probe: 177 trials; and happy incongruent probe: 172 trials.

2.5. Temporospatial PCA

The temporospatial PCA technique was used to identify ERPs elicited to cues and probes in the dot-probe task. PCA belongs to a class of factor-analytic procedures which use eigenvalue decomposition to extract linear combinations of variables (latent factors) in order to account for patterns of covariance in the data parsimoniously (i.e., with the fewest factors) (Dien & Frishkoff, 2005). In ERP data, PCA extracts linear combinations of data from all time points and recording sites to distinguish patterns of electrocortical activity (Kujawa et al., 2013; Pegg et al., 2019).

PCA was conducted separately on the cue and probe data using the ERP PCA Toolkit, version 2.86, in MATLAB (Dien, 2010b). Two ERP averages per subject were entered into the data matrix for the cue PCA (i.e., angry cue, happy cue), and four ERP averages per subject were entered into the data matrix for the probe PCA (i.e., angry congruent probe, angry incongruent probe, happy congruent probe, happy incongruent probe). For both the cue and probe data, a temporal PCA was performed first to separate the ERP components in the temporal domain (Dien, 2010a, 2012; Dien & Frishkoff, 2005). A promax rotation was used, as it is most effective for the temporal PCA (Dien, 2010a, 2012), along with a covariance relationship matrix (Kayser & Tenke, 2003), Kaiser weighting (Dien, Beal, & Berg, 2005), and the kappa for the promax set at 3 (Dien, 2010a). For the decomposition procedure, singular value decomposition was used. The temporal PCA utilized the time points as variables and the subjects, conditions, and recording sites as observations (Dien & Frishkoff, 2005). A parallel test (Horn, 1965) was used on the resulting Scree plot (Cattell, 1966), which compares the Scree of the dataset to that obtained from a fully random dataset. For the cue data, 9 temporal factors accounted for a greater proportion of variance than those generated by the random dataset and accounted for 97.0% of the total variance. For the probe data, 7 temporal factors accounted for a greater proportion of variance than those generated by the random dataset and accounted for 95.6% of the total variance.

These temporal factors were then entered into a spatial PCA (Dien, 2010a, 2012; Dien & Frishkoff, 2005). An infomax rotation was used, as it is most effective for the spatial PCA (Dien, 2010a). The spatial PCA utilized recording sites as variables and the subjects, conditions, and temporal factor scores as observations (Dien & Frishkoff, 2005). The parallel test of this Scree plot extracted 3 spatial factors from each cue temporal factor and 4 spatial factors from each probe temporal factor. Overall, 27 temporospatial factor combinations were generated for the cue dataset and 28 temporospatial factor combinations were generated for the probe dataset.

Temporospatial factors accounting for at least 0.5% of the total variance were subjected to a robust ANOVA (Dien, 2017; Keselman, Wilcox, & Lix, 2003) in the ERP PCA Toolkit to evaluate the effect of emotion (i.e., angry face pairs versus happy face pairs) in the cue condition and the effects of both emotion and congruency (i.e., congruent versus incongruent) in the probe condition. PCA factor scores were converted to microvolt scaling, and 49,999 bootstrapping simulations were run 11 times in order to compute the standard deviation of the resulting p-values; the median p-value was then reported. If twice the standard deviation of the p-values plus the median p-value exceeded the alpha threshold (0.05), the result was treated as a borderline significant result (Dien, 2017). Significant temporospatial factors and their descriptions are reported in Table 1.

Table 1.

Temporospatial factor combinations sensitive to face pair cues and probes.

Factor Combination	Variance (%)	Temporal Peak (ms)	Peak Electrode	Main Effect of Condition TWJt/c (1.0, 22.0) (p)	Description
CUES
TF5/SF3	0.55	38	Oz	Emotion: 6.90 (0.012)	Occipital negativity for happy v. angry face pair cues
TF3/SF1	7.78	86	PO4	Emotion: 4.96 (0.035)	Parieto-occipital positivity for happy v. angry face pair cues
PROBES
TF4/SF3	1.18	220	PO4	Congruency: 5.75 (0.029)	Parieto-occipital positivity for incongruent v. congruent probes

3. Results

3.1. Behavioral (RT) Analyses

Behavioral analyses were conducted in jamovi (Lenth, 2020; R Core Team, 2019; Singmann, 2018; the jamovi project, 2020). Participants made an average of 703.76 hits (SD = 30.79) and 15.08 misses (SD = 30.60) in the dot-probe task. A 2 (emotion: angry versus happy face pair cues) × 2 (congruency: congruent versus incongruent probes) repeated-measures ANOVA was conducted on the RT data. The emotion × congruency interaction was not significant [F(1,24) = 0.224, p = 0.640, η_p² = 0.009]. The main effect of emotion was also not significant [F(1,24) = 0.036, p = 0.851, η_p² = 0.002]. However, the main effect of congruency was significant [F(1,24) = 5.283, p = 0.031, η_p² = 0.180], such that RTs were shorter for congruent (angry-congruent: M = 321.56, SD = 57.72; happy-congruent: M = 321.26, SD = 57.85) versus incongruent (angry-incongruent: M = 323.12, SD = 54.81; happy-incongruent: M = 324.12, SD = 58.99) probes. These results suggest that anxious participants exhibit hypervigilance and greater visual attentional allocation toward emotional (angry and happy) versus neutral face cues. Reaction times to probes as a function of emotion and congruency are shown in Figure 2.

Figure 2.

Reaction times (RTs) to probes as a function of emotion (angry versus happy face pair cues) and congruency (congruent versus incongruent probes) in the dot-probe task. Bars represent standard error of the mean.

3.2. PCA

3.2.1. Cue PCA

Of the 23 factor combinations accounting for more than 0.5% of the total variance, 4 factor combinations were significantly sensitive to emotion (p < .05). Two factor combinations (TF4/SF2 and TF4/SF3) will not be discussed because they had widespread scalp distributions that did not appear to be consistent with commonly observed cue-locked ERPs. However, there was a significant effect of emotion on Temporal Factor 5/Spatial Factor 3 (TF5/SF3), T_WJt/c(1.0, 22.0) = 6.90, p = 0.012, MSe = 0.02, a factor combination consisting of a very early negativity. This factor combination peaks at 38 ms at channel Oz and resembles an early C1 ERP component; the C1 typically onsets 40–60 ms poststimulus and peaks 80–100 ms poststimulus (Luck, 2014). This factor combination presented as an increased negativity for happy compared to angry face pair cues, suggesting that at very early, pre-attentive stages of processing, activity in the primary visual cortex (V1) (Clark & Hillyard, 1996; Eldar et al., 2010; Pourtois, Grandjean, Sander, & Vuilleumier, 2004) is enhanced by happy face pair cues. Following TF5/SF3, there was also a significant effect of emotion on Temporal Factor 3/Spatial Factor 1 (TF3/SF1), T_WJt/c(1.0, 22.0) = 4.96, p = 0.035, MSe = 0.08, a factor combination consisting of an early positivity. This factor combination peaks at 86 ms at channel PO4 and resembles an early P1 component; the P1 typically onsets 60–90 ms poststimulus and peaks 100–130 ms poststimulus (Luck, 2014). This factor combination presented as an increased positivity for happy compared to angry face pair cues, suggesting that more attention is allocated to happy face pair cues at this stage of processing.

3.2.2. Probe PCA

Of the 23 factor combinations accounting for more than 0.5% of the total variance, 4 factor combinations were significantly sensitive to either emotion, congruency, or an interaction between emotion and congruency (p < .05). Three factor combinations (TF1/SF1, TF5/SF1, and TF7/SF2) will not be discussed because they were either noisy or had scalp distributions that did not appear to be consistent with commonly observed probe-locked ERPs. However, there was a significant effect of congruency on TF4/SF3, T_WJt/c(1.0, 22.0) = 5.75, p = 0.029, MSe = 0.06, a factor combination consisting of an early positivity. This factor combination peaks at 220 ms at channel PO4 and resembles a P2 component; the P2 typically peaks around 200 ms poststimulus (Eldar et al., 2010; Helfinstein et al., 2008; Rossignol et al., 2013). This factor combination presented as an increased positivity for incongruent compared to congruent probes, which may reflect more elaborative processing and emotional evaluation of neutral, compared to emotional (angry and happy), faces (i.e., attentional avoidance from the emotional faces). The effect of emotion on TF4/SF3 was nonsignificant, T_WJt/c(1.0, 22.0) = 0.00, p = 0.99, MSe = 0.14, and the interaction between emotion and congruency was also nonsignificant, T_WJt/c(1.0, 22.0) = 2.34, p = 0.16, MSe = 0.08.

The original, grand average cue ERP waveforms at Oz and PO4 and grand average probe ERP waveforms at PO4 are shown in Figure 3, and the ERP waveforms and spatial topographies for the three temporospatial factors are shown in Figure 4.

Figure 3.

ERPs for angry and happy face pair cues at electrode sites Oz and PO4 and probes at electrode site PO4 prior to PCA.

Figure 4.

PCA temporospatial factor ERPs and scalp distributions for TF5/SF3 (C1-Cue), TF3/SF1 (P1-Cue), and TF4/SF3 (P2-Probe). Temporal peaks are indicated with dashed lines on the ERP waveform figures and peak electrodes are indicated with black circles on the scalp distribution figures.

3.3. Split-Half Reliability

Using the PCA results to inform ERP scoring, we sought to determine whether the mean value around the PCA-derived C1-Cue, P1-Cue, and P2-Probe peaks were reliably measured in the original ERP waveforms. First, for both the cue and probe datasets, odd and even trials were averaged separately. Temporal peak and peak electrode information from the PCA (see Table 1) were used to determine appropriate time windows and locations to search for peaks in the ERP waveforms. A 28–48 ms search window at Oz was used to identify the C1-Cue peak, a 76–96 ms search window at PO4 was used to locate the P1-Cue peak, and a 210–230 ms search window at PO4 was used to identify the P2-Probe peak. The mean value around the peaks (50 ms) for the odd and even average waveforms were then exported from Brain Vision Analyzer. Finally, Spearman-Brown-corrected correlations were computed on the odd and even averages to assess split-half reliability.

Reliability coefficients suggested strong split-half reliability for the C1 to angry cues (Spearman r = 0.92, p < .001), C1 to happy cues (Spearman r = 0.87, p < .001), P1 to angry cues (Spearman r = 0.96, p < .001), P1 to happy cues (Spearman r = 0.85, p < .001), P2 to angry congruent probes (Spearman r = 0.81, p < .001), P2 to angry incongruent probes (Spearman r = 0.82, p < .001), P2 to happy congruent probes (Spearman r = 0.85, p < .001), and P2 to happy incongruent probes (Spearman r = 0.70, p < .001). These results provide additional insight into the peaks for the three components and suggest that the mean value around the peaks tends to be reliably measured in the ERP waveforms.

3.4. Secondary Exploratory Analysis: The Effect of Emotional Face Position

A secondary exploratory analysis was conducted to determine the extent to which ERPs derived through PCA were moderated by the position of the emotional (angry or happy) face. As before, a PCA was conducted on the cue data; however, four averages per subject were entered into the data matrix for this PCA (i.e., angry-neutral (angry face on the left), neutral-angry (angry face on the right), happy-neutral (happy face on the left), and neutral-happy (happy face on the right)). Temporospatial factors accounting for at least 0.5% of the total variance were subjected to a robust ANOVA to evaluate the effects of both emotion and position (i.e., emotional face on the left versus emotional face on the right).

Of the 29 factor combinations accounting for more than 0.5% of the total variance, no factor combinations were significantly sensitive to the interaction between emotion and position (p < .05). However, four factor combinations were significantly sensitive to emotion, and one of these factor combinations (TF3/SF1), T_WJt/c(1.0,22.0) = 4.85, p = 0.049, MSe = 0.24, again resembled an early P1 component, peaking 86 ms poststimulus at channel PO8 and presenting as an increased positivity for happy compared to angry face pair cues. Additionally, one factor combination (TF6/SF3) was significantly sensitive to position, T_WJt/c(1.0,22.0) = 6.05, p = 0.024, MSe = 0.11, peaking 178 ms poststimulus at channel P4. This factor combination showed decreased amplitudes for right-sided (ipsilateral to P4) emotional faces (neutral-angry and neutral-happy) compared to left-sided (contralateral to P4) emotional faces (angry-neutral and happy-neutral). The remaining factor combinations (TF1/SF2, TF4/SF2, TF4/SF3) will not be discussed because they were noisy or had scalp distributions that did not appear to be consistent with commonly observed cue-locked ERPs.

4. Discussion

The present analyses are part of a larger, ongoing study investigating the neurobiological mechanisms of MBCT, including ERP markers of threat-related attentional bias in adults with moderate to high levels of anxiety using a dot-probe task adapted from Mueller et al. (2009). The PCA analyses were performed on the dot-probe task ERP data administered prior to the MBCT intervention. The goal of this preliminary study was to utilize PCA to systematically identify the timing and scalp distributions of ERPs elicited to cues and probes in the aforementioned dot-probe task in an anxious adult population. Temporospatial PCA identified 2 components sensitive to face pair cues. The first cue-locked component was an early relative negativity for happy versus angry face pair cues peaking around 38 ms poststimulus over central occipital sites. This component appears to be consistent with the C1-Cue observed in previous dot-probe studies with anxious and healthy adults (Eldar et al., 2010; Pourtois et al., 2004). The second cue-locked component was an early relative positivity for happy versus angry face pair cues peaking around 86 ms poststimulus over parieto-occipital sites. This component appears to be consistent with the P1-Cue previously observed in dot-probe studies with anxious adults (Helfinstein et al., 2008; Mueller et al., 2009; Rossignol et al., 2013). Temporospatial PCA also identified one component sensitive to incongruent versus congruent probes. The probe-locked component was an early relative positivity that was enhanced for incongruent (probes replacing neutral faces) compared to congruent (probes replacing emotional faces) probes peaking around 220 ms poststimulus over parieto-occipital sites. This component appears to be consistent with the P2-Probe observed in a previous dot-probe study with healthy adults (Pintzinger, Pfabigan, Pfau, Kryspin-Exner, & Lamm, 2017). We subsequently used the PCA results to determine whether the mean value around the PCA-derived C1-Cue, P1-Cue, and P2-Probe peaks were reliably measured in the original ERP waveforms. Reliability was strong for all three components, suggesting that the mean value around these peaks tends to be reliably measured in the ERP waveforms.

Gupta et al. (2019) conducted a review on the neural chronometry of threat-related attentional bias which showed that early ERP components, including the P1, N170, P2, and N2pc, are modulated by threatening and emotional stimuli in anxious populations, reflecting enhanced allocation of attention to threat and emotion at earlier stages of processing. In the present study, we also observed emotion and congruency-related modulations of early components in anxious adults performing the dot-probe task, but our findings were somewhat surprising. We observed that the C1-Cue component peaked around 38 ms poststimulus and was enhanced for happy versus angry face pair cues, suggesting enhanced, pre-attentive processing of happy faces, or early perceptual-level avoidance of angry faces, at the level of V1. In a previous dot-probe study comparing anxious and nonanxious participants, between-group analyses indicated that anxious participants had a more negative C1 amplitude compared to nonanxious participants in response to angry-neutral face pairs, but the two groups did not differ in their C1 amplitudes in response to happy-neutral face pairs (Eldar et al., 2010). Similarly, in another dot-probe study focusing on healthy adults, within-group analyses indicated that the C1 component was enhanced for fearful, compared to happy, faces (Pourtois et al., 2004). Additionally, in both of the aforementioned studies, the C1 peaked later (~80–90 ms poststimulus) than the C1-Cue identified in the present study. Studies have suggested that C1 modulation by the cue’s emotional valence in the dot-probe task could result from interactions between V1 and subcortical limbic structures (e.g., the amygdala) responsible for threat detection (Eldar et al., 2010; Pourtois et al., 2004). Thus, the enhanced C1 to happy face pair cues in our study may reflect a very early form of avoidance from the threatening (angry) face pair cues. Additionally, our C1-Cue results suggest that anxious adults differentiate emotion conditions at very early stages of information processing, further supporting the view that ERPs are a promising method to examine the time course of attentional bias at the neural level.

We also observed that the P1-Cue component peaked around 86 ms poststimulus and was enhanced for happy compared to angry face pair cues, suggesting that relatively more attention was allocated to the happy face pair cues. Interestingly, in previous dot-probe studies with anxious adults, the P1 peaked later (~100–150 ms poststimulus) (Helfinstein et al., 2008; Mueller et al., 2009; Rossignol et al., 2013). Additionally, prior work has indicated that the P1 is enhanced to angry-neutral face pairs in adults with social anxiety, but it is unclear whether this finding extends to other types of anxiety. For example, between-group analyses from Helfinstein et al. (2008) demonstrate that high socially anxious individuals displayed higher mean P1 amplitudes to angry-neutral face pairs compared to low socially anxious individuals, suggesting increased sensory processing of faces in individuals with high levels of social anxiety. Within-group analyses conducted by Mueller et al. (2009) similarly showed that individuals with SAD display larger P1 amplitudes to angry-neutral versus happy-neutral face pairs, suggesting an early hypervigilance to angry faces. However, between-group analyses conducted by Rossignol et al. (2013) demonstrated that, compared to low socially anxious individuals, high socially anxious individuals displayed increased P1 amplitudes in response to neutral-emotional face pairs (neutral–angry, neutral–happy, neutral–disgust and neutral–fear), irrespective of the emotional expression included in the pair. These results suggest that a generalized hypervigilance to emotional faces may occur in social anxiety. On the other hand, our P1-Cue results suggest the possibility that participants high in trait anxiety may tend to avoid the angry face pair cues at early stages of processing (Gupta et al., 2019).

Finally, we observed that P2-Probe amplitudes were larger for incongruent versus congruent probes. These results may reflect more elaborative processing and emotional evaluation of neutral, compared to emotional (angry and happy), faces (i.e., attentional avoidance from the emotional faces). Interestingly, in a previous dot-probe study with healthy adults, within-group analyses showed that women displayed larger P2 amplitudes in congruent compared to incongruent negative conditions, which could indicate sustained attentional engagement with negative, compared to neutral, information (Pintzinger et al., 2017). The vigilance-avoidance model suggests that anxious individuals initially direct attentional resources toward threat during early, automatic stages of processing, but then direct their attention away from threat during later, more strategic stages of processing in an attempt to avoid detailed elaborative processing of threatening material (Mogg, Bradley, De Bono, & Painter, 1997; Mogg, Mathews, & Weinman, 1987). Our results instead suggest that anxious adults display avoidance from emotional (both angry and happy) face stimuli.

The rapid nature of the dot-probe design used in the present study may explain the earlier C1-Cue and P1-Cue latencies observed in our results. We utilized a 100 ms cue presentation duration, but many of the studies referenced herein utilized longer durations for cue stimuli, such as 500 ms (Eldar et al., 2010; Helfinstein et al., 2008; Pintzinger et al., 2017; Rossignol et al., 2013). We are particularly interested in examining early forms of attentional bias; thus, the brief cue duration was well-suited to our investigations. The rapid nature of the task may also cause ERP components to overlap. Indeed, both our cue and probe PCAs revealed that there are several distinct components emerging in a very short period of time (300 ms poststimulus), highlighting the complexity of these neural responses. PCA is a powerful tool which allows us to extract components which may have otherwise been masked in the typical ERP results. For example, the P1-Cue and P2-Probe components are not evident in the original ERP cue and probe waveforms at PO4, likely due to the overlap of several components in the waveforms (see Figure 3). However, PCA helps disentangle these overlapping components so we can clearly observe the P1-Cue and P2-Probe (see Figure 4).

A secondary exploratory analysis was conducted to test the extent to which ERPs derived through PCA were moderated by the position of the emotional (angry or happy) face. No temporospatial factors were sensitive to the interaction between emotion and position. However, one component was significantly sensitive to position, peaking 178 ms poststimulus at channel P4. This factor combination showed decreased amplitudes for right-sided (ipsilateral to P4) emotional faces (neutral-angry and neutral-happy) compared to left-sided (contralateral to P4) emotional faces (angry-neutral and happy-neutral). The timing and electrode position of this temporospatial factor is somewhat similar to that of an N2pc component, which typically occurs 200–300 ms poststimulus at posterior scalp sites contralateral to an attended object (Luck, 2014). However, the ipsilateral versus contralateral effect was surprising and did not suggest that anxious participants attend more to emotional, versus neutral, faces, as observed in prior studies specifically designed to elicit the N2pc. Prior dot-probe studies have shown that high-trait anxious, socially anxious, and healthy populations display N2pc amplitude or latency modulations to threatening and other emotional faces, reflecting rapid orienting to threatening and emotional cues, in general (Gupta et al., 2019). However, it is important to note that the dot-probe task used in the present study was adapted from Mueller et al. (2009), and this prior study primarily focused on examining the P1 component elicited to face pair cues and probes and was not specifically designed to elicit the N2pc component. In the current version of the task, the fixation cross was presented centrally on the lower part of the screen, and face cues were presented in the upper visual field. However, several dot-probe studies designed to elicit the N2pc component present the fixation cross at the center of the screen with cues presented to the left and right of this central fixation cross (Fox et al., 2008; Holmes, Bradley, Kragh Nielsen, & Mogg, 2009; Kappenman et al., 2014, 2015; Reutter, Hewig, Wieser, & Osinsky, 2017). Future studies aiming to investigate the N2pc will benefit from a design that can better test the differences between spatial orientation of the cues relative to the fixation cross followed by application of temporospatial PCA to isolate the relevant components.

Behavioral analyses indicated that RTs were shorter for congruent versus incongruent probes, suggesting that anxious participants exhibit hypervigilance and greater visual attentional allocation toward emotional (angry and happy) versus neutral face cues. Interestingly, the P2-Probe PCA results appeared to reflect more elaborative processing and emotional evaluation of neutral, compared to emotional (angry and happy), faces (i.e., attentional avoidance from the emotional faces). The enhanced P2 amplitude elicited to probes appearing in the location of neutral, versus emotional, face cues suggests that incongruent probes more immediately capture attention because participants are already attending to the location of the neutral face. However, the RT effects suggest that participants are attending more toward the location of the emotional face at the time of probe onset, and are therefore faster to respond to congruent probes. Therefore, another possibility is that the enhanced P2 amplitude elicited to incongruent probes reflects the greater amount of attentional resources required when shifting attention toward incongruent probes. Additionally, these differing results further reinforce the importance of including neurophysiological measures (i.e., ERPs) along with (often unreliable) behavioral measures to improve the metrics obtained from the dot-probe task.

This preliminary study has a number of limitations which should be acknowledged. First, the present study lacks a healthy comparison group; thus, it is unclear whether the PCA-derived C1-Cue, P1-Cue, and P2-Probe components are specifically anxiety related or whether these components would also be observed in non-anxious samples. However, these three components are particularly promising and reliable ERP measures of attentional bias that can be extended to designs comparing anxious and healthy comparison groups.

Secondly, in order to determine whether our findings varied as a function of anxiety, we performed exploratory correlations between the STAI-T scores, the temporospatial factor peaks (TF5SF3, TF3SF1, TF4SF3), and the mean amplitudes around the PCA-derived peaks in the ERP waveforms (C1-Cue, P1-Cue, P2-Probe). None of the amplitude values were significantly correlated with STAI-T scores. While these results were surprising, the lack of correlation between trait anxiety and ERPs may result from the fairly restricted range of STAI-T scores across participants. Specifically, all participants had moderate to high levels of anxiety, indexed by a score of 40 or above on the STAI-T. Although well-powered to detect within-subjects condition effects, the relatively small sample size in this preliminary study may also have limited our ability to detect more modest between-subjects effects of trait anxiety on ERPs.

The third limitation arises from a lack of correction for multiple comparisons. In order to account for potential Type I errors resulting from multiple comparisons, a false discovery rate (FDR) correction, such as the Benjamini-Hochberg method (Benjamini & Hochberg, 1995), is often applied to the robust ANOVAs. However, the purpose of this study was to use PCA to identify cue- and probe-locked components, their scoring windows, and their electrode sites in our dot-probe task in order to inform hypotheses for our larger, ongoing study. Thus, we wished to minimize Type II errors which would have likely arisen from a stringent FDR correction. Therefore, in the present study, if the robust ANOVA indicated that there was a significant (p < .05) effect of emotion in the cue condition or emotion and/or congruency in the probe condition, we retained the temporospatial factor for further evaluation. Fortunately, there are several strengths associated with using robust ANOVAs. The robust statistics function generates inferential statistical tests comparable to ANOVAs that are designed to be more robust against violations of statistical assumptions. This robust statistic features trimmed means and winsorized variances/covariances to minimize effects of outliers, a bootstrapping routine to estimate the sample mean distribution rather than making the assumption that the data is normally distributed, and a Welch-James approximate degrees of freedom statistic (resulting sometimes in decimal degrees of freedom) that avoids the assumption of homogeneous error variances/covariances. The latter also makes it unnecessary to use Greenhouse-Geisser or Huynh-Feldt epsilon corrections since sphericity is not assumed (Dien, 2017; Keselman et al., 2003). Overall, however, replication of this study is still required due to the lack of correction for multiple comparisons.

A final limitation pertains to the face cue stimuli used in the present study. In accordance with Mueller et al. (2009), the present study utilized faces from Ekman’s Pictures of Facial Affect (Ekman & Friesen, 1976). All of the happy faces used in the present study exhibited smiles with exposed teeth, while half of the angry faces used featured exposed teeth and the other half featured compressed lips. Although faces were matched as closely as possible, it is possible that differences in the mouth regions of the angry faces differentially affected the ERPs. It has been shown that larger P1, N170, vertex positive potential (VPP), and slow positive wave (SPW) ERPs occur to mouth expressions with teeth, and that high luminance/contrast in the mouth-teeth border may drive these ERP effects (DaSilva et al., 2016). Indeed, early visual components are also sensitive to low-level stimulus features such as luminance and contrast (Johannes, Münte, Heinze, & Mangun, 1995). Fortunately, there is some evidence that luminance levels of the Ekman faces may not differ significantly. Pourtois et al. (2004) conducted a dot-probe experiment similar to the one used in the present study using Ekman faces. After face stimuli were trimmed to exclude hair and non-facial contours, there were no significant differences in low-level properties (e.g., luminance, spatial frequency) for the different emotional face conditions. In order to further investigate the effects of exposed and non-exposed teeth on ERP components, future studies may implement a brief passive viewing task featuring emotional faces with and without exposed teeth prior to selecting facial stimuli for implementation in attentional bias tasks. While the aforementioned limitation makes it difficult to interpret the emotion-related effects of our PCA-derived components, our results still highlight that the C1-Cue, P1-Cue, and P2-Probe are promising and reliable markers to investigate in future attentional bias studies.

5. Conclusion

The present study is the first, to our knowledge, to use temporospatial PCA to systematically identify ERP components sensitive to emotionally-valenced face pair cues and spatially-relevant probes in a dot-probe task in adults with moderate to high levels of anxiety. Results highlight three reliably elicited components that are of interest for future research. One factor combination resembled a C1-Cue, consisting of an early negativity at 38 ms poststimulus over central occipital sites. The component was enhanced for happy versus angry face pair cues, suggesting that enhanced, pre-attentive processing of happy faces and avoidance of angry faces occurs at the level of V1. The subsequent factor combination resembled a P1-Cue, consisting of an early positivity at 86 ms over parieto-occipital sites. The component was also enhanced for happy versus angry face pair cues, indicating enhanced allocation of attention to the happy faces and avoidance from the angry faces. The final factor combination resembled a P2-Probe, consisting of an early positivity at 220 ms poststimulus over parieto-occipital sites. The component was enhanced for incongruent compared to congruent probes, which may reflect more elaborative processing and emotional evaluation of neutral, compared to emotional (angry and happy), faces (i.e., attentional avoidance from the emotional faces). These results highlight the C1-Cue, P1-Cue, and P2-Probe as promising and reliable neurophysiological markers for attentional bias research and suggest that anxious adults display avoidance from angry face stimuli. It is recommended that future ERP attentional bias studies utilize PCA to systematically identify the timing and scalp distribution of ERPs elicited to task-related stimuli.