Distress Intolerance Moderation of Attention to Emotion: An Eye-Tracking Study

Abstract

Distress intolerance (DI) is an important individual difference reflective of the inability to endure aversive affective states and is relevant to multiple clinical populations, but underlying emotional processing mechanisms remain unclear. The current study used eye-tracking to examine biased attention towards emotional stimuli at baseline and in the context of acute stress in a non-clinical sample (N = 165). We hypothesized that DI would incrementally predict greater stressor-elicited increases in sustained/delayed disengagement, but not initial orientation/facilitated engagement negative (i.e., threat, dysphoric) attention biases, and that DI’s association with maladaptive stress regulation would depend on these increases. Partially consistent with predictions, DI was only independently associated with stressor-elicited increases in sustained negative bias and, unexpectedly, decreases in sustained positive bias. Further, DI and change in sustained threat bias marginally interacted to predict cardiovascular but not subjective anxious mood recovery. Theoretical implications are discussed.

Introduction

Distress intolerance (DI), defined as the inability to withstand aversive affective states, is an individual difference thought to contribute to multiple forms of psychopathology (Leyro et al. 2010). For instance, DI has been linked to greater risk of relapse/treatment dropout in individuals with substance use disorders (Brown et al. 2009; Daughters et al. 2005; Strong et al. 2012) and greater symptom severity in individuals diagnosed with mood/anxiety conditions (Allan et al. 2014; Macatee et al. 2015, 2013). DI appears to be an important person-level moderator of response to negative affect with high relevance to psychopathology; however, theoretical understanding of DI is limited and its underlying mechanisms remain unclear.

DI is conceptualized as a trait-like variable related to but distinct from the propensity to experience negative affect (Leyro et al. 2010). Empirical findings have largely supported this conceptualization. Self-report (i.e., perceived ability to withstand aversive affective states) and behavioral (i.e., actual ability to persist on a distressing laboratory task) measures of DI have demonstrated temporal stability (Cummings et al. 2013; Kiselica et al. 2014; Macatee et al. 2015) and predictive validity (e.g., stress-elicited relapse, perseverative thought symptoms) independent of negative affect (Brown et al. 2009; Cougle et al. 2011; Macatee et al. 2013, 2015; Strong et al. 2012; Volz et al. 2014). One plausible mechanism underlying individual differences in DI is attention bias towards negative emotional stimuli, defined as selective attentional allocation to threatening and dysphoric relative to neutral stimuli (Gibb et al. 2016). Attention bias towards negative emotional information is not a unitary phenomenon, but varies across time (e.g., initial orientation vs sustained attention towards negative emotional stimuli) and in relation to task demands (e.g., facilitated engagement with goal-relevant negative emotional stimuli vs slowed disengagement from goal-irrelevant negative emotional stimuli) (Armstrong and Olatunji 2012; Cisler and Koster 2010). Sustained attention towards and delayed disengagement from negative stimuli refers to the degree to which such stimuli capture and hold attention, whereas initial orientation and facilitated engagement refers to the speed at which attention orients towards negative emotional stimuli (Armstrong and Olatunji 2012; Cisler and Koster 2010). With respect to DI, there is theoretical and empirical support for a relationship between DI and sustained attention/delayed disengagement attention biases specifically.

Given the centrality of negative appraisals of distress and its increased motivational relevance to the latent DI construct (McHugh and Otto 2012), it may be that, during acute distress states, negative emotional information becomes more salient to individuals with high DI and attracts greater sustained cognitive processing relative to non-negatively valenced information. Greater sustained processing of incidental negative emotional information during acute stress (e.g., unpleasant interoceptive sensations/cognitions) in conjunction with deficits in disengaging attention from such information may amplify avoidance action urges (e.g., substance use) and interfere with the processing of goal-relevant neutral information necessary for adaptive resolution of the stressor. Thus, increased sustained attention towards and difficulties disengaging from negative emotional stimuli during acute stress, referred to as stress-elicited sustained negative attention bias (NAB) and stress-elicited delayed disengagement NAB throughout the remainder of this manuscript, may be reciprocally related to negative appraisals of distress and avoidance urges, ultimately impairing goal-directed persistence behavior in aversive contexts and maintaining DI over time. Although no studies known to the authors have explicitly examined the relationship between DI and different forms of NAB, two studies have found positive relations between persistence on a stressful laboratory task and the ability to control attention in the presence of emotional stimuli (Johnson 2009a, b). However, these studies did not utilize validated measures of DI (e.g., McHugh and Otto 2011, 2012) nor did they employ methodology capable of directly assessing different forms of attention bias (e.g., eye-tracking; Sanchez et al. 2013). Further, these investigations were not able to measure the relationship between DI and stress-elicited increases in sustained/delayed disengagement NAB, an important consideration given extant experimental and longitudinal data on DI’s relationship with maladaptive behavior in the context of acute distress.

A number of longitudinal studies suggest that DI’s relationship with maladaptive behavior is dependent upon the presence of acute stress (Abrantes et al. 2008; Macatee et al. 2013; Strong et al. 2012; Volz et al. 2014). In a sample of smokers making a quit attempt, Abrantes et al. (2008) found that increases in negative affect on quit day predicted higher risk for early lapse for high but not low DI individuals. Similar to Abrantes et al. (2008) findings, Strong et al. (2012) found that greater DI predicted higher likelihood of opioid use in patients initiating buprenorphine treatment, with the effect particularly strong early after the first dose and during titration, periods where negative affect is expected to be heightened (Eder et al. 1998; Lintzeris et al. 2002). In another study of smokers attempting to quit, Volz et al. (2014) did not find a significant main effect of DI on daily craving, though they did find a positive relationship between DI and daily craving on days with greater stressors. Similarly, Macatee et al. (2013) found a non-significant main effect of DI on daily obsessions, though there was a trend-level positive relationship between the two in the context of greater daily stressors.

Existing experimental studies utilizing mood inductions have also supported the context dependence of DI’s relationship with maladaptive behavior (Norberg et al. 2015; Perkins et al. 2010; Shaw and Timpano 2016). Perkins et al. (2010) found a significant positive relationship between DI and smoking behavior in healthy adult daily smokers after overnight abstinence but not after a neutral mood induction. Similarly, Norberg et al. (2015) found that greater DI predicted decreased discarding behavior after a sad mood induction in a non-clinical sample, but was unrelated to discarding after a neutral mood induction. In a related hoarding study, Shaw and Timpano (2016) found that DI was positively related to saving behavior after a stress but not a neutral mood induction in a non-clinical sample. Taken together, extant longitudinal and experimental data suggest that DI’s relationship to multiple forms of maladaptive behavior is moderated by acute distress, a notion consistent with DI’s theoretical conceptualization as a moderator of response to negative affect distinct from negative affect itself (Leyro et al. 2010). Thus, if attention bias towards negative emotional information is a clinically-relevant mechanism underlying individual differences in DI, DI should specifically predict increases in sustained attention/delayed disengagement NAB under conditions of acute negative affect. Further, if stress-elicited sustained/delayed disengagement NAB is an important mechanism underlying DI’s relationship to maladaptive stressor responding, DI should predict such responding for individuals who demonstrate stress-elicited increases in sustained attention/delayed disengagement NAB.

Few studies have examined DI’s relationship with the time course of emotional responding to stressors, though some work has suggested that DI may specifically predict impaired affective recovery from stressors as opposed to greater initial reactivity to them (Cougle et al. 2011; Macatee and Cougle 2015). Cougle et al. (2011) found that, after controlling for general negative affect, DI significantly predicted greater anxious mood persistence but not immediate anxious reactivity to instructed visualization of a personally-relevant distressing thought. Using the same paradigm, Macatee and Cougle (2015) found that, compared to those on the waitlist, individuals who received a DI intervention reported significantly lower urges to neutralize the thought 2 min following visualization but not immediately afterwards, again suggesting that DI may contribute to maladaptive responding to stressors via impairments in emotional recovery. Interestingly, some studies have also found a more robust relationship between NAB and impaired emotional recovery relative to greater emotional reactivity (Clasen et al. 2013; Sanchez et al. 2013). It is plausible that DI’s relationship with impaired stressor recovery is dependent upon stressor-elicited sustained/delayed disengagement NAB, though to our knowledge this possibility has never been explicitly tested.

The purpose of the present study was to test the hypothesized association between DI and stress-elicited NAB using eye-tracking methodology. Compared to reaction time measures of attention bias, eye-tracking methods produce more psychometrically reliable indices (Price et al. 2015) and allow for the measurement of multiple forms of attention bias. Sanchez et al. (2013) developed an eye-tracking task capable of distinguishing multiple forms of attention bias relevant to our hypotheses (i.e., initial orientation, sustained attention, facilitated engagement, and delayed disengagement); this task was used in the present study. First, we hypothesized that (1) DI would moderate the effect of a stressor (i.e., anticipation of uncertain social-evaluative threat) on change in sustained and delayed disengagement attentional bias towards negative (i.e., threatening and dysphoric affective stimuli), but not positive, emotional information such that individuals with high DI would demonstrate increases in sustained/delayed disengagement NAB, but not initial orientation or facilitated engagement NAB, from pre to post- stressor. Moreover, we predicted that the effect of DI on stressor-elicited change in sustained/delayed disengagement NAB would be unaccounted for by the magnitude of negative emotional reactivity to the stressor, trait social-evaluative anxiety symptoms, or past week depressive symptoms. Including these covariates in the model will allow for a theoretically-driven test of the hypothesis that DI’s relationship with stressor-elicited NAB is not attributable to stressor-specific (i.e., emotional reactivity to the stressor/trait social-evaluative anxiety) or general negative affect (i.e., past week depressive symptoms). We also hypothesized that (2) DI and stressor-elicited increases in sustained attention/delayed disengagement NAB would significantly interact to predict decreased emotional recovery from the stressor, but would be unrelated to initial reactivity to the stressor. Specifically, we predicted that greater DI would predict worse negative mood recovery, as indexed by self-reported negative affect and cardiovascular activation, only in the context of greater stressor-elicited increases in sustained/delayed disengagement NAB.

Method

Participants

To test study hypotheses, a non-clinical sample recruited for a larger investigation on DI (Macatee et al. 2016) and related emotion dysregulation indices (i.e., resting HRV, threat bias and worry; Macatee et al. 2017a, b) was utilized (N = 165; 77% female; M age = 19.32, SD = 1.96; Caucasian: 70.9%, Hispanic: 13.3%, African-American: 11.5%, Asian: 1.2%, American Indian or Alaskan Native: 0.6%, Other: 2.4%). Primary analyses¹ presented in the current manuscript have not been published elsewhere. To ensure adequate variability in DI, individuals in the top quartile of DI scores as assessed by self-report (DII; McHugh and Otto 2012) were oversampled through targeted e-mail invitations to individuals scoring in the top 20% of DII scores on a separate screening questionnaire completed by undergraduate psychology students for course credit. A young adult non-clinical sample was deemed to be an appropriate sample to test our hypotheses for the following reasons: (1) A non-clinical sample was thought to be less susceptible to ceiling effects with respect to emotional reactivity to the laboratory stressor compared to a clinical sample; (2) Young adults, as opposed to adolescents or children, were recruited given existing data demonstrating self-reported DI’s relative stability in this population (Kiselica et al. 2014). Participants earned course credit for completing the study.

Measures

Questionnaires

Distress Intolerance

The Distress Intolerance Index (DII; McHugh and Otto 2012) is a 10-item self-report questionnaire designed to index an individual’s perceived ability to tolerate distressing states (e.g., “I can’t handle feeling distressed or upset”). Most importantly, the DII has demonstrated superior convergent validity with behavioral measures of DI relative to other affect tolerance measures (McHugh and Otto 2011; McHugh et al. 2016; Seo and Kwon 2016; Szuhany and Otto 2015; but see; Williams et al. 2015). In the current study, the DII demonstrated excellent internal consistency (M = 10.44, SD = 9.19; α = 0.94).

Social‑Evaluative Anxiety

The Social Interaction Anxiety Scale (SIAS; Mattick and Clarke 1998) is a 20-item self-report measure of social anxiety that has demonstrated good internal consistency and test–retest reliability in previous literature (Mattick and Clarke 1998). In the current study, the SIAS demonstrated excellent internal consistency (M = 19.55, SD = 13.99; α = 0.94).

General Distress—Depressive Symptoms

The general distress—depressive (GDD) symptoms subscale of the Mood and Anxiety Symptom Questionnaire-Short Form (MASQ: Keogh and Reidy 2000) is a 12-item scale that was used to assess past week depressive symptoms in the present study. Although the MASQ-GDD is designed to assess depressive symptoms and does show some specificity to individuals with clinical depression, the items load best onto a general distress factor (Keogh and Reidy 2000). Thus, the MASQ-GDD is well-suited to testing the hypothesized moderating role of DI on stress-elicited NAB independent of general distress symptoms. In the current sample, the MASQ-GDD demonstrated excellent internal consistency (M = 27.60, SD = 10.94; α = 0.94).

State Mood

Three visual analog scales (VASs) for each mood state (happy, anxious, sad) were used to assess state mood at three different times during the experiment. Each scale consisted of an emotion word and a line with 11 anchor points ranging from 0 (not at all) to 10 (very much); participants were asked to indicate the extent to which they were experiencing that particular emotion at that moment. Each mood state was assessed using three emotion words (happy mood: optimistic, joyful, happy; anxious mood: nervous, tense, anxious; sad mood: upset, sad, depressed). Mean ICs were excellent for each mood state (happy: α = 0.95; anxious: α = 0.93; sad: α = 0.92; see Table 1 for descriptives).

Table 1.

Change in affect across experiment

	Time 1: pre-stress M (SD)	Time 2: anticipatory stress M (SD)	Time 3: post-stress M (SD)
Self-report
Happy	16.27 (7.11)	16.69 (7.51)	16.43 (7.74)
Depressed	4.71 (5.69)	4.23 (5.42)	3.11 (4.36)***
Anxious	7.47 (7.19)	10.43 (8.49)***	6.03 (6.53)***
Psychophysiology
Mean HR	75.10 (9.30)	79.31 (10.33)***	76.50 (9.94)***

^***Significant difference from previous assessment at p < .001

HR heart rate. Reproduced from Macatee et al. 2017a, 2017b

Attention Task

Stimuli Materials

Stimuli were pairs of face images consisting of an emotional and neutral facial expression of the same person. The images were derived from the Karolinska Directed Emotional Faces (KDEF) database (Lundqvist et al. 1998) and used by Sanchez et al. (2013). Images were modified to fit into an oval window, and the hair, neck, and surrounding parts of the pictures were darkened to remove irrelevant aspects of the faces. KDEF images were chosen based upon emotional intensity and prototypicality of the corresponding facial emotion (see Sanchez and Vazquez 2013 for validation data). The final stimuli set consisted of 36 happy, angry, and sad facial expressions (18 men and 18 women for each emotion), together with a corresponding neutral expression by the same person.

Experimental Set‑Up and Attention Indices

The attention task consisted of six practice trials and 108 experimental trials in which an emotional and neutral face were presented on opposite sides of the screen (36 happy—neutral, 36 sad—neutral, 36 angry—neutral). Trial order was randomized and emotional and neutral faces were presented equally often on the left and right sides of the screen. To replicate the stimuli size/position ratios used by Sanchez et al. (2013), stimuli were presented on a 41 cm (width) × 30.5 cm (height) screen. Each face was 9 cm (width) × 12.7 cm (height). Facial stimuli were centered on the screen, 18.1 cm apart (measured from their centers). Participants were seated approximately 90 cm from the screen’s center, resulting in a visual angle of approximately 5.7 degrees between each picture’s center and the screen’s center. The attention task design used in the present study was developed by Sanchez et al. (2013) and has also been described by Macatee et al. (2017a). Each trial started with a blank black screen for 500 ms, followed by a centered white fixation cross for 500 ms. A random number then appeared on the center of the screen for 1000 ms. Participants were instructed to fixate on the number and silently name it as quickly as possible to ensure the participant’s gaze is focused on the screen prior to facial stimuli presentation; note that participants did not say the number out loud because of concerns that vocalization would affect the positioning of the participant’s head in the mount. After the offset of the random number, an emotion-neutral face pair (happy—neutral, sad—neutral, angry—neutral) was presented for 3000 ms during which time participants were instructed to look at the screen freely (i.e., the naturalistic viewing period). Fixation data recorded during the 3000 ms period was used to calculate two indices of naturalistic emotional information processing (e.g., Kellough et al. 2008) for each emotion-neutral category: (1) Initial orientation—the probability of initially fixating on an emotional face after onset of the facial pair—was derived by dividing the number of trials in which the participant initially fixated upon the emotional face by the total number of trials for that emotion-neutral category; and (2) Sustained attention—the proportion of dwell time on the emotion face relative to the neutral face across the 3000 ms naturalistic viewing period—was derived by grand averaging the proportion of dwell time indices across all trials for a particular emotion-neutral category.

The engagement-disengagement portion of the attention task began after the 3000 ms naturalistic processing period and allowed for an assessment of both the ability to disengage from and engage with emotional stimuli. The engagement-disengagement task consisted of three conditions: (1) On one-third of the trials in each emotion-neutral category, attentional engagement with the emotional face was assessed. After 3000 ms of naturalistic processing, the ‘wait for fixation period’ began during which the task awaited a fixation of at least 100 ms on the neutral face. As soon as the 100 ms fixation was registered, the ‘wait for fixation period’ ended and a frame (square or circle) appeared surrounding the emotional face. Participants were instructed to move their gaze as quickly as possible toward the framed face and to identify the frame shape on the keyboard using one of two response keys. Thus, engagement trials allowed for an assessment of how long it took participants to disengage from the neutral face and fixate on the emotional face. (2) On another one-third of the trials in each emotion-neutral category, attentional disengagement from the emotional face was assessed. After 3000 ms of naturalistic processing, the ‘wait for fixation period’ began during which the task awaited a fixation of at least 100 ms on the emotional face. As soon as the 100 ms fixation was registered, the ‘wait for fixation period’ ended and a frame (square or circle) appeared surrounding the neutral face. As with the engagement condition trials, participants were instructed to move their gaze as quickly as possible toward the framed face and to identify the frame shape on the keyboard using one of two response keys. Thus, disengagement trials allowed for an assessment of how long it took participants to disengage from the emotional face and fixate on the neutral face. (3) The remaining one-third of trials in each emotion-neutral category consisted of control trials in which a new trial began immediately after the 3000 ms naturalistic processing period. Disengagement, engagement, and control condition trials for each emotion-neutral category were randomly presented for each participant, with both types of frame shapes equally likely to appear on both sides of the screen for each condition.

In line with Sanchez et al. (2013) procedure, fixation data collected after the end of the ‘wait for fixation period’ were used to compute engagement and disengagement latencies for each emotion-neutral category. Engagement latency was computed by adding 100 ms (i.e., the neutral face fixation duration required to end the ‘wait for fixation period’) to the amount of time that elapsed from the end of the ‘wait for fixation period’ to the first 100 ms fixation on the framed, emotional face. Disengagement latency was computed by adding 100 ms (i.e., the emotional face fixation duration required to end the ‘wait for fixation period’) to the amount of time that elapsed from the end of the ‘wait for fixation period’ to the first 100 ms fixation on the framed, neutral face.

Eye‑Tracking Device

The desktop mounted EyeLink 1000 (SR Research Ltd, Ottawa, Ontario, Canada) was used to record participants’ eye movements. The EyeLink 1000 is a video camera-based infrared eye-tracking system that uses a velocity-based event-detection algorithm to filter raw gaze samples into saccade, fixation, and blink events; velocity thresholds for saccade detection were set to the recommended values for cognitive research (30 degrees/second; EyeLink 1000 User Manual, Version 1.5.2). Gaze data were acquired from the right eye at a sampling rate of 1000 Hz. A 9-point calibration and validation procedure was conducted prior to starting the attention task to configure the system such that the spatial accuracy error was below 0.5 degrees on average. Further, calibration accuracy was re-checked after every trial. Participants were required to fixate on a central fixation point; if error was greater than 1 degree, the calibration and validation procedure was conducted again to obtain a spatial accuracy of less than 0.5 degrees of average error. The participant’s head was kept stable using an adjustable head mount, with a distance between the participant’s eyes and the camera of approximately 60 cm. All participants had normal to corrected-to-normal vision and were allowed to wear contact lenses or glasses during the attention tasks if necessary; further, participants were required to remove eye make-up before completing the attention tasks.

Stress Induction

Speech Anticipation Task

To assess attentional bias under stress, a speech anticipation procedure developed by Waugh, Panage, Mendes, and Gotlib (2010) was utilized. After participants completed the baseline attention task, they rested for 5 min (i.e., the pre-stress period) and then rated their current mood. The experimenter then placed a camera in the room and told participants that they would have 2 min to prepare a 5-min speech during which they would be recorded and judged by evaluators on their clarity, coherence, and persuasiveness. The speech topic was “Why are you a good friend?”, a topic that has been used successfully in prior studies to elicit anticipatory stress responses (Sanchez et al. 2013). Participants were told that there would be two coin flips that would determine if and when the participants gave the speech. They were informed that the first coin would be flipped immediately after the 2 min speech preparation period and would determine if they gave the speech immediately, or waited for the second coin flip after the second attention task, which would determine whether they gave the speech at that time or not at all. The experimenter then informed the participant to remain still and silent so as not to disturb the heart rate recording and left the participant alone for 2 min to prepare their speech (i.e., the anticipatory-stress period). After the 2 min speech preparation period, the first coin was flipped and participants were told that a computer randomizer would determine if the result of the coin flip indicated that they would give the speech now or wait until the second attention task was completed. All participants were told that they would not give the speech immediately and that the second coin flip, taking place after the second attention task, would determine whether they gave the speech. After participants rated their mood, they completed the second attention task and then the second coin flip was conducted. All participants were told that they would not be giving the speech, and they rested for 5 min (i.e., the post-stress recovery period) before completing a final mood rating.

Waugh et al. (2010) speech anticipation procedure was chosen for the following reasons: (1) it was important to assess attention bias within a stressful context rather than after stressor cessation (e.g., after giving a speech) to test the hypothesis that DI would moderate emotional information processing during acute stress, (2) unpredictable threat is relevant to DI-linked clinical phenomena (e.g., worry, substance use problems; Gorka et al. 2016; Hogle et al. 2010; Macatee et al. 2015, 2016), and (3) the speech anticipation procedure has been used in a prior eye-tracking study examining relations between mood recovery and attention bias indices (Sanchez et al. 2013).

Psychophysiology Recording

Polar RS800CX

Participants wore heart rate monitors (i.e., Polar Electro RS800CX wristwatch monitors; Anderson and Hope 2009) to record heart rate at a sampling rate of 1000 Hz from the pre-stress relaxation period through the end of the post-stress recovery period. The heart rate monitor consisted of a wristwatch and a dampened two-lead elastic band worn underneath the clothing around the sternum, and has been used in prior work on cardiovascular stress reactivity (Gouin et al. 2014). Timestamps were recorded to mark the beginning of the 5 min pre-stress period, the 2 min anticipatory-stress period, and the 5 min post-stress period.

Procedure

All procedures were approved by Florida State University’s institutional review board and were in accordance with the ethical standards of the institutional research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. After providing informed consent, participants completed questionnaires and the baseline attention task. Next, the heart rate monitor was attached and participants sat quietly for a 5 min pre-stress relaxation period, after which they rated their current mood. Participants then completed the speech-anticipation period, rated their current mood, and completed the attention task again. After completing the post-stress attention task, participants sat quietly for a 5 min post-stress recovery period, after which they rated their current mood. Finally, participants completed six online questionnaire assessments over the following two weeks which were not relevant to the current study aims but are reported on elsewhere (Macatee et al. 2016, 2017a).

Results

Data Preprocessing

Attention Task

Naturalistic Attention Bias

Of the 165 participants consented, 137 completed the baseline attention task and 126 completed the post-stress attention task (see Macatee et al. 2017a for more information on reasons for exclusion). An independent sample t-test revealed no significant differences in DI between completers and non-completers, t(163) = −0.30, p = .77. Automated scripts written in Perl were developed to parse the event data ASCII files generated by the EyeLink system. The EyeLink generates a paired saccade- blink event to demarcate recording periods during which the pupil could not be found (e.g., due to a blink or technical error). Paired saccade-blink events that occurred during the free-viewing portion of a trial were identified and removed by automated parsing scripts written in Perl (% of total free-viewing samples removed for baseline and post-stress tasks, respectively; M = 8.0, SD = 6.6, range: 0.3–33.4; M = 9.8, SD = 6.6, range: 0.4–31.5). Prior work has used a 40% criterion for participant exclusion due to excessive signal loss (Graham et al. 2011); thus, no participants were excluded for this reason. Next, individual trials were removed if the participant made a saccade away from the center of the screen before facial stimuli onset or if a fixation was not made on either facial stimulus during the 3000 ms free-viewing period (1.6% and 2.2% of all trials in the baseline and post-stress tasks, respectively). One outlier had 26.9 and 64.8% of their trials removed on the baseline and post-stress attention tasks, respectively; this participant was excluded from all naturalistic attention bias analyses. Finally, fixations on either facial stimulus during the 3000 ms free-viewing period with durations ≥ 100 ms were extracted. In line with recent recommendations (Gibb et al. 2016), split-half reliabilities were computed for initial orientation and sustained attention bias across the three emotion-neutral categories for the baseline and post-stress attention tasks (see Table 2).

Table 2.

Descriptives for attention bias indices

	Baseline			Stress
	M	SD	Split-half reliability	M	SD	Split-half reliability
Initial orientation
Happy	0.58	0.08	−0.03(n = 133,31 trials) −0.17(n = 94,36 trials)	0.60	0.10	0.30(n = 125,31 trials) −0.11(n = 85,36 trials)
Angry	0.50	0.07	−0.49(n = 133,32 trials) to −0.46(n = 93,36 trials)	0.50	0.07	0.01(n = 125,30 trials) to −0.31(n = 76,36 trials)
Sad	0.49	0.06	−0.63(n = 133,30 trials) to −1.07(n = 95,36 trials)	0.50	0.08	−0.27(n = 125, 28 trials) to −0.03(n = 91,36 trials)
Sustained attention
Happy	0.57	0.06	0.70(n = 133,31 trials) −0.76(n = 94,36 trials)	0.57	0.07	0.76(n = 125,31 trials) −0.79(n = 85,36 trials)
Angry	0.53	0.05	0.56(n = 133,32 trials) −0.58(n = 93,36 trials)	0.52	0.05	0.67(n = 125,30 trials) −0.61(n = 76,36 trials)
Sad	0.51	0.05	0.44(n = 133,30 trials) −0.54(n = 95,36 trials)	0.52	0.05	0.51(n = 125, 28 trials) −0.66(n = 91,36 trials)
Engagement
Happy	469.84	55.97	0.66(n = 131,5 trials) −0.79(n = 42,12 trials)	469.80	57.20	0.43(n = 123,5 trials) −0.86(n = 38,12 trials)
Angry	480.50	64.31	0.62(n = 132,5 trials) −0.75(n = 48,12 trials)	476.60	66.24	0.52(n = 123,5 trials) −0.81(n = 43,12 trials)
Sad	478.64	63.49	0.67(n = 129,5 trials) −0.68(n = 43,12 trials)	467.93	55.41	0.67(n = 123,5 trials) −0.82(n = 44,12 trials)
Disengagement
Happy	469.98	60.95	0.75(n = 131,5 trials) −0.71(n = 44,12 trials)	466.81	53.23	0.77(n = 123,6 trials) −0.88(n = 39,12 trials)
Angry	465.03	52.95	0.58(n = 132,5 trials) −0.76(n = 46,12 trials)	465.45	51.70	0.57(n = 121,5 trials) −0.69(n = 51,12 trials)
Sad	471.18	63.53	0.68(n = 131,5 trials) −0.56(n = 41,12 trials)	467.71	54.10	0.61(n = 124,6 trials) −0.89(n = 42,12 trials)

Initial orientation/sustained attention means are proportions. Engagement/Disengagement means are latencies in milliseconds; data fofr a given engagement/disengagement measure was excluded if they had less than three SD valid trials available. Split-half reliabilities were computed using the entire sample as well as a subsample with all valid trials for each naturalistic and engagement/disengagement index of attention bias. Reproduced from Macatee et al. 2017a

Engagement/Disengagement Attention Bias

Automated perl scripts were developed to derive engagement and disengagement latencies from the EyeLink’s ASCII event output files according to Sanchez et al. (2013) procedure. The following criteria were used to ensure that the engagement/disengagement latencies reflected valid shifts in gaze elicited by the appearance of the shape (86.3 and 87.1% of baseline and post-stress trials, respectively): (1) Gaze was on the non-framed facial stimulus for at least 100 additional ms after the end of the ‘wait for fixation period’ and appearance of the frame; (2) Fixation on the framed facial stimulus lasted for at least 100 ms; (3) Time elapsed from the end of the ‘wait for fixation period’ to the first 100 ms fixation on the framed face was <1000 ms. Further, trials were considered invalid if the participant incorrectly identified the shape of the frame. Seven outliers (i.e., defined as <3 SD valid trials for at least one emotion category) were identified in the baseline task data (valid trials: M = 10.55; SD = 1.59) and five outliers were identified in the post-stress task data (valid trials: M = 10.65; SD = 1.45); outliers in a particular emotion category were excluded from those analyses.

Heart Rate

Of the 137 participants with at least one valid attention task data point, 91, 94, and 108 had physiological data for the pre-stress, anticipatory-stress, and post-stress periods, respectively; missing data occurred primarily because of technical errors and transmission signal loss during the experiment. Heart rate data were cleaned and extracted using a five step process. First, we used the Polar Pro Trainer 5TM software to export the collected IBI data to a text file for preprocessing before analysis. Second, the IBI series was loaded into HRV analysis software (HRVAS; Ramshur 2010) and visually inspected for ectopic intervals. Third, an automated filter was utilized to confirm visually identified ectopic intervals by marking IBIs that differed more than 20% from the previous interval; these intervals were then corrected using cubic spline interpolation (% of IBIs interpolated; pre-stress: M = 2.0, SD = 2.7, range: 0–14.9; anticipatory-stress: M = 2.0, SD = 3.6, range: 0–30; post-stress: M = 2.8, SD = 3.9, range 0–24.2). Fourth, the corrected IBI series was detrended using the discrete wavelet packet transform (Shafqat et al. 2007). Lastly, HRVAS was used to compute mean HR for the pre-stress, anticipatory- stress, and post-stress periods.

Data Analytic Strategy

To account for the lack of independence inherent to repeated measurements, multilevel marginal models were constructed to test hypotheses. Although repeated measures ANOVA analyses are commonly used for these kind of data, only multilevel models can accommodate missing within-subject observations and model correlated within-subject residuals without imposing assumptions (e.g., sphericity, compound symmetry) (Quené and van der Bergh 2004). These advantages were particularly relevant to the present study given the prevalence of missing data for attention bias, state mood, and mean HR. Thus, multilevel marginal models were used for all analyses to maximize inclusion of all valid observations. Further, in line with recent recommendations (Gurka et al. 2011), no assumptions were imposed on the within-subject covariances and they were modeled using an unstructured matrix.

Of the 136 subjects with at least one valid attention task data point for naturalistic bias, 11 did not have post-stress attention task data, whereas three subjects with post-stress data did not have baseline data. For engagement/disengagement bias, 137 subjects had at least one valid attention task data point, with 12 of these subjects only having baseline data and three only having post-stress data. Among the available valid data for naturalistic and engagement/disengagement biases, 12 participants’ anxious mood scores were missing at the pre-stress and/or anticipatory-stress time points due to participant error, precluding computation of an anxious mood reactivity score, leaving 124 subjects available for testing hypothesis one for naturalistic attention biases and 125 subjects for engagement/disengagement biases. Of the 122 subjects with valid baseline and post-stress naturalistic attention task data, 107 had at least one mean HR data point and all 122 had at least one state mood data point, leaving 107 and 122 subjects available to test hypothesis two for sustained attention bias for mean HR and anxious mood, respectively. Of the 117 subjects with valid baseline and post-stress disengagement data for each emotion, 103 had at least one mean HR data point and all 117 subjects had at least one state mood data point, leaving 103 and 117 subjects available to test hypothesis two for delayed disengagement bias for mean HR and anxious mood, respectively.

To probe significant within-subject main effects, post-hoc comparisons with Bonferroni corrections were utilized. Significant within-subject interactions were probed by decomposing the effect into separate models with one within-subject factor; significant omnibus tests were then probed using Bonferroni-corrected follow-up comparisons. To probe significant interactions between within-subject factors and between-subject variables, follow-up regression analyses were conducted in which change scores were entered as the dependent variable. Change scores were computed according to Burt and Obradovic’s (2013) guidelines for evaluating the relative reliability of residualized change vs. difference scores. Burt and Obradovic (2013) recommend using residualized change scores when the correlation between the baseline and post measurement is smaller than their corresponding variance ratio, whereas difference scores are recommended when the correlation is greater than the variance ratio. Further, the authors recommend a difference score be used when there is a large correlation between the baseline measurement and the outcome variable. These guidelines were used when computing the change score to be used as the outcome variable in follow-up regression analyses for significant between-subject*within-subject interactions.

Statistical Analyses

Speech Anticipation Stressor Manipulation Check

Descriptives for state mood and mean HR at each time point are presented in Table 1.

In the first model, the effect of the speech anticipation stressor on self-reported negative affect was evaluated. State mood was entered as the dependent variable. Emotion was entered as a three-level within-subject factor (happy mood, anxious mood, dysphoric mood) and Time was also entered as a three-level within-subject factor (pre-stress, anticipatory-stress, post- stress). Both factors were entered as fixed effects (see Online Resource 1 for more detail). As expected, anxious mood significantly increased from pre-stress to anticipatory-stress period and decreased from the anticipatory-stress to post-stress recovery period. In contrast, depressed mood was significantly greater during the anticipatory-stress relative to post-stress period whereas depressed mood did not significantly change from the pre-stress to anticipatory-stress period. No change in happy mood was observed throughout the experiment. Because only self-reported anxious, but not depressed or happy, mood significantly changed from the pre-stress to anticipatory-stress period, only this index of self-reported mood change across the experiment was used in subsequent analyses.

In the second model, the effect of the speech anticipation stressor on mean HR was evaluated. Mean HR was entered as the dependent variable. Time was entered as a three-level within-subject factor (pre-stress, anticipatory-stress, post-stress) fixed effect. Results revealed a significant effect of Time, F(2, 94.07) = 21.27, p < .001; Bonferroni-corrected follow-up comparisons revealed that mean HR was significantly greater during the anticipatory-stress relative to the pre-stress, t(94.02) = 6.51, p < .001, and post-stress, t(95.11) = 4.48, p < .001, periods.

The results demonstrate that the uncertain speech anticipation manipulation successfully elicited a subjective and physiological stress response that subsequently decreased after the possible speech threat was removed.

DI and Change in Attention Bias towards Emotional Stimuli under Acute Stress

Descriptives and reliabilities for attention bias indices are presented in Table 2. Attention bias (i.e., initial orientation, sustained attention, facilitated engagement, delayed disengagement) was entered as the dependent variable to test the role of DI on attention bias in four separate omnibus tests. Emotion was entered as a three-level within-subject factor (positive, threat, dysphoric) and Time was entered as a two-level within-subject factor (baseline, stress). DI, anxious mood reactivity to the stressor, trait social-evaluative anxiety, and past week depressive symptoms were entered as between-subject predictors. Anxious mood reactivity to the stressor was computed using a residualized change score (i.e., regressing anticipatory-stress anxious mood on pre-stress anxious mood and saving the standardized residuals); this change score was chosen based upon Burt and Obradovic’s (2013) guidelines (see Online Resource 1 for more detail). All between-subject predictor*within-subject factor interactions were included in the model. All variables were entered as fixed effects.

Initial Orientation

Please see Online Resource 1 for examination of Emotion, Time, and Emotion*Time effects. As hypothesized, the DI*Emotion*Time interaction was non-significant, F(2, 115.76) = 0.54, p = .58. All between-subject main effects were non- significant, ps > 0.08, as were all other between-subject*within-subject interactions, ps > 0.11.

Sustained Attention

Please see Online Resource 1 for examination of Emotion, Time, and Emotion*Time effects. All between-subject main effects were non-significant, ps > 0.27. Interactions between the main effect of Emotion and trait social anxiety, past week depressive symptoms, and state anxious reactivity were non-significant, ps > 0.75, whereas a marginal DI*Emotion interaction was observed, F(2, 115.27) = 3.03, p = .052^.2 Interactions between the main effect of Time and DI, trait social anxiety, and past week depressive symptoms were non-significant, ps > 0.24, though the Time*anxious reactivity interaction was significant, F(1, 112.40) = 6.85, p = .01. As hypothesized, the DI*Emotion*Time interaction was significant, F(2, 112.87) = 5.06, p = .008 (see Fig. 1). Interestingly, the social anxiety*Emotion*Time interaction was also significant, F(2, 113.06) = 5.14, p = .007, whereas the past week depressive symptoms*Emotion*Time and anxious reactivity*Emotion*Time interactions were non-significant, ps > 0.57.

Fig. 1.

Estimated marginal means for sustained attention bias at baseline and under stress for positive (top panel), threat (middle panel), and dysphoric (bottom panel) emotions at high (+1 SD) and low (−1 SD) levels of DI

To probe the significant interactions, three hierarchical regression analyses were conducted to examine the relationship between all between-subject variables (i.e., DI, trait social anxiety, past week depressive symptoms, state anxious reactivity) and stressor-elicited change in sustained attention bias for each emotion (i.e., positive, threat, dysphoric) (see Table 3). Trait social anxiety, past week depressive symptoms, and state anxious reactivity were entered in step one of the model. In order to assess the incremental variance in stressor-elicited change in attention bias explained by DI, DI was entered in step two of the model. Residualized change scores were created for each attention bias variable by regressing sustained attention bias under acute stress on baseline attention bias and saving the standardized residuals as an index of change. These changes scores were chosen based upon Burt and Obradovic’s (2013) guidelines (see Online Resource 1 for more information).

Table 3.

Effect of distress intolerance on change in sustained attentional bias

		ΔR2	ΔF	β	T	P
ΔSustained positive attentional bias
Step 1		0.04	1.60			0.194
	Stressor reactivity			−0.192	−1.994	0.049
	SIAS total			0.140	1.203	0.231
	MASQ total			−0.128	−1.132	0.260
Step 2		0.04	4.13			0.045
	Stressor reactivity			−0.210	−2.205	0.030
	SIAS total			0.207	1.736	0.085
	MASQ total			0.006	0.047	0.963
	DII total			−0.259	−2.032	0.045
ΔSustained threat attentional bias
Step 1		0.06	2.455			0.067
	Stressor reactivity			−0.158	−1.658	0.100
	SIAS total			−0.188	−1.642	0.103
	MASQ total			0.130	1.160	0.249
Step 2		0.07	8.856			0.004
	Stressor reactivity			−0.132	−1.427	0.156
	SIAS total			−0.284	−2.459	0.015
	MASQ total			−0.061	−0.482	0.631
	DII total			0.368	2.976	0.004
ΔSustained dysphoric attentional bias
Step 1		0.07	2.743			0.047
	Stressor reactivity			−0.155	−1.634	0.105
	SIAS total			−0.026	−0.228	0.820
	MASQ total			0.222	1.990	0.049
Step 2		0.04	4.355			0.039
	Stressor reactivity			−0.136	−1.454	0.149
	SIAS total			−0.094	−0.801	0.425
	MASQ total			0.086	0.676	0.501
	DII total			0.262	2.087	0.039

DII Distress Intolerance Index, SIAS Social Interaction Anxiety Scale, MASQ Mood and anxiety symptom questionnaire-depression subscale, Stressor Reactivity self-reported anxious mood reactivity

Contrary to hypotheses, inclusion of DI in step two significantly improved the model for change in sustained attention towards positive stimuli (i.e., ΔSustainedPositiveBias), ΔF(1, 110) = 4.13, p = .045, ΔR2 = 0.04, revealing a significant inverse relationship between DI and ΔSustainedPositiveBias, β = −0.26, t(110) = −2.03, p = .045. As predicted, inclusion of DI in step two significantly improved the model for change in sustained attention towards threat stimuli (i.e., ΔSustainedThreatBias), ΔF(1, 110) = 8.86, p = .004, ΔR2 = 0.07, revealing a significant positive relationship between DI and ΔSustainedThreatBias, β = 0.37, t(110) = 2.98, p = .004. As hypothesized, inclusion of DI in step two significantly improved the model for change in sustained attention towards dysphoric stimuli (i.e., ΔSustainedDysphoricBias), ΔF(1, 110) = 4.36, p = .04, ΔR2 = 0.04, revealing a significant positive relationship between DI and ΔSustainedDysphoricBias, β = 0.26, t(110) = 2.09, p = .04.

To probe the significant state anxious mood reactivity*Time interaction, the regression coefficients for anxious mood reactivity across all three analyses were inspected, revealing an overall relationship between greater mood reactivity to the stressor and larger stressor-elicited decreases in sustained attention towards all emotional stimuli. Due to the significant trait social anxiety*Emotion*Time interaction, the differential relationship between social anxiety and ΔSustainedPositiveBias, ΔSustainedThreatBias, and ΔSustainedDysphoricBias was inspected; the regression analyses revealed a significant, inverse relationship between social anxiety and ΔThreatBias, β = −0.28, t(110) = −2.46, p = .02, but not ΔSustainedPositiveBias or ΔSustainedDysphoricBias, ps > 0.08, suggesting that greater social anxiety was related to larger stressor-elicited decreases in sustained attention towards threat stimuli only.

An exploratory analysis to determine if the significant effect of DI on stressor-elicited change in sustained attention bias varied over the three second time course revealed a non-significant interaction (see Online Resource 1).

Facilitated Engagement

The main effects of Emotion and Time as well as the Emotion*Time interaction were non-significant, ps > 0.13. As hypothesized, the DI*Emotion*Time interaction was non- significant, F(2, 112.64) = 0.72, p = .49. All other between-subjects main effects, ps > 0.09, as well as interactions between within-subject factors and between-subjects predictors were non- significant, ps > 0.08.

Delayed Disengagement

The main effects of Emotion and Time as well as the Emotion*Time interaction were non-significant, ps > 0.35. Contrary to hypotheses, the DI*Emotion*Time interaction was non- significant, F(2, 108.79) = 0.10, p = .91. All other between-subjects main effects, ps > 0.11, as well as interactions between within-subject factors and between-subjects predictors were non-significant, ps > 0.07.

DI, Stressor‑Elicited Change in Attention Bias towards Emotional Stimuli, and Emotional Reactivity and Recovery

Anxious mood and mean HR were entered as the dependent variables in two separate omnibus tests of DI and change in sustained attention bias’s association with emotional reactivity and recovery. Time was entered as a three-level within-subject factor (pre-stress, anticipatory- stress, post-stress). DI, stressor-elicited change in attention bias for each emotion type, and their interactions were entered as between-subject predictors. Further, all between-subject predictor*within-subject factor interactions were included in the model. All variables were entered as fixed effects. Because DI was non-significantly related to delayed disengagement bias, the planned analyses involving delayed disengagement and emotional reactivity/recovery, which revealed null interactions between DI, ΔDisengagementBias, and Time, have been moved to Online Resource 1.

Subjective Anxious Mood

Contrary to predictions, the DI*ΔSustainedThreatBias*Time, F(2, 112.19) = 0.03, p = .97, and DI*ΔSustainedDysphoricBias*Time, F(2, 112.97) = 0.22, p = .81, interactions were non- significant. The DI*ΔSustainedPositiveBias*Time was also non-significant, F(2, 111.77) = 0.31, p = .73. The main effect of DI was significant, F(1, 114.66) = 56.93, p < .001, which revealed a positive relationship between DI and anxious mood throughout the experiment. All other between-subject and between-subject*within-subject interaction effects were non-significant, ps > 0.15.

Mean Heart Rate

As expected, the DI*ΔSustainedPositiveBias*Time interaction was non-significant, F(2, 78.11) = 0.83, p = .44. Contrary to hypotheses, the DI*ΔSustainedDysphoricBias*Time interaction was also non-significant, F(2, 78.47) = 0.43, p = .65. Partially consistent with predictions, a marginally significant DI*ΔSustainedThreatBias*Time interaction emerged, F(2, 79.61) = 3.06, p = .052. All other between-subject and between-subject*within-subject interaction effects were non-significant, ps > 0.10. Visualization of the marginal interaction suggested an effect consistent with hypotheses (see Online Resource 2), and so follow-up regression analyses were conducted.

Two hierarchical regression analyses were conducted to examine the interactive effect between DI and ΔSustainedThreatBias on mean HR reactivity to and recovery from the laboratory stressor. The main effect of DI was entered in step one, and the main effects of ΔSustainedThreatBias, ΔSustainedPositiveBias, and ΔSustainedDysphoricBias were entered in step two. The DI/ΔSustainedPositiveBias and DI/ΔSustainedDysphoricBias interaction terms were entered in step three. In order to assess the incremental variance explained by the DI/ΔSustainedThreatBias interaction term, it was entered in step four of the model. Residualized change scores for mean HR reactivity and mean HR recovery were computed by regressing mean HR at the subsequent measurement point on the prior measurement point and saving the standardized residuals as an index of change (i.e., anticipatory-stress mean HR regressed on pre- stress HR for HR reactivity; post-stress recovery mean HR regressed on anticipatory-stress mean HR for HR recovery). Thus, more positive HR reactivity scores indicate greater HR reactivity to the stressor and more positive HR recovery scores indicate worse HR recovery. Residualized change scores were chosen based upon Burt and Obradovic’s (2013) guidelines (see Online Resource 1 for more information).

As expected, inclusion of the DI/ΔSustainedThreatBias interaction term in step four did not significantly improve the model, ΔF(1, 77) = 0.05, p = .83, ΔR2 = 0.001, for mean HR reactivity. In contrast and consistent with hypotheses, inclusion of the DI/ΔSustainedThreatBias interaction in step four did significantly improve the model, ΔF(1, 78) = 5.33, p = .024, ΔR2 = 0.054, for mean HR recovery. To probe the interaction, the relationship between DI and HR recovery was examined at increased (+1 SD) and decreased (−1 SD) ΔSustainedThreatBias. As predicted, at increased ΔSustainedThreatBias greater DI significantly predicted worse HR recovery, β = 0.48, t(78) = 3.00, p = .004, whereas the relationship was non-significant at decreased ΔSustainedThreatBias, β = −0.14, t(78) = −0.73, p = .47.

Discussion

Overall, the results of the present study were partially consistent with hypotheses. As hypothesized, DI predicted increased sustained, but not initial orientation attentional bias towards threatening and dysphoric emotional information in the context of anticipation of uncertain social-evaluative threat. Moreover, DI predicted increased sustained NAB in the context of acute stress independent of theoretically-relevant covariates, suggesting that DI’s association with stressor-elicited sustained NAB is not a function of stressor-specific or general negative affect. In contrast to these findings and inconsistent with hypotheses, DI also predicted decreased sustained attentional bias towards positive emotional information in the context of uncertain social-evaluative threat anticipation. Although unexpected, this finding also remained significant after controlling for stressor-specific and general negative affect. Further, DI was non-significantly related to baseline sustained attention bias, suggesting that DI is associated with altered attentional processing of emotional information specifically during acute stress rather than with trait-level emotional processing biases. In contrast to the sustained attention results and our hypotheses, DI was non-significantly related to stressor-elicited delayed disengagement NAB, suggesting that DI is linked to preferential processing of negative emotional stimuli under stress but is unrelated to efficiency of goal-directed disengagement from these stimuli when prompted by a task-relevant cue.

Partial support was found for the impaired emotional recovery hypothesis. Contrary to hypotheses, DI did not interact with stress-elicited sustained or delayed disengagement NAB to predict impaired subjective anxious mood recovery; rather, only a significant main effect of DI emerged, with greater DI associated with higher anxious mood ratings across all time points. In contrast, though the omnibus test did not reach significance, the pattern of results for mean HR was partially consistent with hypotheses such that DI interacted with stressor-elicited sustained attention bias towards threat in the prediction of cardiovascular recovery but not reactivity. However, DI did not significantly interact with stress-elicited delayed disengagement NAB to predict cardiovascular recovery.

The present study’s results provide strong empirical support for DI’s relationship with stress-elicited sustained NAB. Given that depressive symptoms, social anxiety symptoms, and state emotional reactivity have all been linked to NAB in prior work (Armstrong and Olatunji 2012; Nelson et al. 2015), our test of DI’s hypothesized role in modulating emotional information processing in the context of a stressor was particularly stringent. In fact, examination of the pattern of associations between the covariates and stressor- elicited NAB revealed that the variables most specific to the stressor itself (i.e., anxious mood reactivity; trait social-evaluative anxiety symptoms) generally predicted decreases in NAB, particularly for threat, which may reflect threat avoidance related to the nature of the social- evaluative stressor. Overall, the results suggest that DI’s association with increased sustained attention towards task-irrelevant, negatively-valenced emotional information in the context of a stressor does not appear to reflect state-level negative affect or stressor-related concerns. Moreover, stressor-elicited change in sustained attention towards threat may be an important moderator of the relationship between DI and physiological recovery from stressors, though the marginally significant omnibus test and discrepant results across physiological and subjective stressor recovery suggest that future research is required before firmer conclusions can be drawn.

Given the centrality of stress-elicited avoidance behavior to the DI construct, the observed robust association between DI and stress-elicited sustained attention towards rather than away from negative emotional cues may appear counterintuitive. One possibility is that high DI individuals demonstrate a vigilance-avoidance attentional pattern, where negative emotional stimuli are initially attended but avoided over time, as has been found in multiple studies of anxious individuals (Armstrong and Olatunji 2012). However, these studies largely used similar or shorter stimulus presentation times to those used in the present study, suggesting that the naturalistic viewing period was sufficiently long for a vigilance-avoidance pattern to emerge. Instead, null associations were observed with initial orientation NAB, and DI’s relationship with stress-elicited sustained NAB did not significantly vary across the three second naturalistic viewing period. Nevertheless, future studies should consider longer stimulus presentation times to directly test the possibility that DI is related to attentional avoidance at later (i.e., >3000 ms) time windows. Another possible explanation for the counterintuitive relationship between DI and sustained NAB is that avoidance behavior in high DI individuals may be a consequence of difficulties diverting attention away from negatively-valenced emotional information under stress, resulting in reliance on more extreme forms of escape behavior (e.g., self-harm, substance use). However, the non-significant relationship between DI and stress-elicited delayed disengagement NAB observed in the current study is not supportive of this possibility, at least for overt visual disengagement from negative emotional cues; future studies should explicitly measure covert attentional disengagement from negative emotional stimuli to further explore this possibility. Finally, it is possible that high DI individuals demonstrate increased sustained attention towards rather than away from negative emotional cues under acute stress because this behavior has been historically reinforced in similar contexts, possibly through the principle of second-order reinforcement (Premack 1959). In other words, attending to negative emotional cues under stress is reinforced for high DI individuals because they enable the relief derived from subsequent escape behavior. To test this account, future studies should consider measuring stress-elicited attention biases towards negative as well as relief (e.g., substances) cues to determine if they are positively associated for high but not low DI individuals.

Contrary to our hypotheses, a significant, inverse association between DI and stressor- elicited positive attentional bias (PAB) emerged. This finding suggests that individuals with high DI, in addition to demonstrating increased NAB under stress, also demonstrated decreased PAB. Although unexpected, the association between DI and positive emotional information processing is consistent with a recent study that found positive affect-related variables to be significantly linked with multiple behavioral measures of DI (Kiselica et al. 2015). Further, increased ability to volitionally direct attention away from threatening and towards positive stimuli under stress has been found to predict greater persistence on a distressing laboratory task (Johnson 2009b). Decreased processing of positive emotional information in the context of stress may contribute to DI via decreasing the salience of temporally delayed positive reinforcers typically paired with persistence behavior (e.g., quitting cigarettes to enhance long-term health; Trafton and Gifford 2011). Decreased processing of positive emotional information may also contribute to DI by reducing the accessibility of alternative, benign interpretations of negative emotional experiences. Indeed, greater efficiency at switching attention from the neutral to the emotional aspects of positively-valenced stimuli has been linked with increased ability to utilize cognitive reappraisal in response to a laboratory stressor (Malooly et al. 2013). Future work could test these possibilities by measuring PAB in the context of a laboratory stressor and assessing its relationship to indices of in-vivo reappraisal ability and delayed reinforcer salience during distressing persistence tasks.

There are a number of limitations of the present investigation. First, a non-stress induction control condition was not utilized, leaving open the possibility that bias changes from pre to post-stressor could be due to other factors (e.g., fatigue), though the significant relationship between anxious mood reactivity to the stressor and bias changes argues against this possibility. Nevertheless, future studies should utilize a control condition to strengthen conclusions regarding causal associations. Second, only a self-report indicator of DI was used. Although the DII was explicitly chosen due to accumulating evidence of its convergence with behavioral DI measures (McHugh and Otto 2011; McHugh et al. 2016; Seo and Kwon 2016; Szuhany and Otto 2015), the observed associations have generally been small to medium in magnitude and one study failed to find statistically significant convergence (Williams et al. 2015). Thus, the behavioral measures, which assess the latency to quit a distressing laboratory task, capture unique variance in the DI construct and should be included (e.g., Strong et al. 2003) in future studies investigating the DI and NAB relationship. Third, although observed associations were robust to multiple theoretically-relevant covariates, effect sizes were generally modest; however, this is not uncommon for correlations across units of analysis (e.g., Patrick et al. 2013). Further, large associations between assessments of general traits (i.e., self-reported DI) and a “snapshot” of behavior in a specific lab context (i.e., attentional bias in the context of a stressor) should not be expected to share large amounts of variance (e.g., see Cyders and Coskunpinar 2011). Given that behavioral measures of DI assess persistence duration on a distressing laboratory task and thus provide a similar singular measurement of behavior, they may demonstrate larger associations with stressor-elicited NAB. Fourth, reliabilities for initial orientation indices were poor, which has also been reported in other eye-tracking studies (Waechter et al. 2014), limiting strong conclusions regarding the role of early NAB in DI. Future studies should consider alternative measures of early attentional orienting. Fifth, heart rate data was inadequately recorded for a relatively large number of subjects, resulting in prevalent missing data; future studies with long recording periods should ensure heart rate signal transmission is adequate and stable. Sixth, although the stress induction utilized in the present study elicited physiological arousal and subjective feelings of anxiety, it is unclear if similar attention biases would emerge in the context of other commonly used laboratory mood inductions. Relatedly, it is unclear if similar biases would be observed in tasks using verbal rather than facial affective stimuli. Future studies should investigate the relative efficacy of different mood inductions for eliciting NAB with facial vs. verbal affective stimuli in individuals with high DI. Finally, it is possible that the null effects observed for some of the moderation analyses are attributable to inadequate statistical power; future studies should recruit larger samples to rule out this possibility.

Supplementary Material

Electronic supplementary material The online version of this article (doi:10.1007/s10608-017-9869-2) contains supplementary material, which is available to authorized users.