Predictive Attentional Bias Modification Induces Stimulus-Evoked Attentional Bias for Threat

Attentional biases are automatic processes that influence the selection of information for further processing (Cisler & Koster, 2010; Kane & Engle, 2003). Spatial attentional biases, involving the direction of attention to the location of salient cues, can be measured via dot-probe or visual probe tasks (VPTs) (MacLeod, Mathews, & Tata, 1986). Faster responses to probe stimuli appearing at the location of certain cue categories provides an implicit measure of bias (Cisler & Koster, 2010; Mogg & Bradley, 2016; Notebaert, Crombez, Van Damme, De Houwer, & Theeuwes, 2011). An interesting application of research into attentional biases is their use in training paradigms, termed Attentional Bias Modification, ABM (MacLeod & Mathews, 2012). ABM aims to reverse a putatively harmful attentional bias in order to change associated behavior or symptoms, such as spider phobia (Luo et al., 2015), depression (Ferrari, Möbius, van Opdorp, Becker, & Rinck, 2016; Wells & Beevers, 2010), addiction (Schoenmakers et al., 2010) and PTSD (Kuckertz et al., 2014). Opinions are strongly divided, however, on the efficacy of ABM, and it appears that, at the least, its efficacy is conditional on moderating factors (Clarke, Notebaert, & MacLeod, 2014; Cristea, Kok, & Cuijpers, 2015; Gladwin, Wiers, & Wiers, 2016).

Some recent studies have raised interesting possibilities potentially relevant to ABM. First, cued Visual Probe Tasks (cVPTs) have been developed to the aim of studying outcome-related attentional biases (Gladwin, 2016). In the cVPT, instead of presenting salient cues and determining how they affect attention, visually neutral predictive cues are used. The predictive value is caused by Picture trials, in which the predictive cues are replaced by an actual exemplar from the category associated with them, for example, threat versus neutral. Assessment of the bias is based on Probe trials, on which instead of the exemplar a probe stimulus requiring a response is presented. Thus, performance is not dependent on a given trial’s specific exemplars, but on the predicted categories of stimuli that could have been presented. Possibly partly due to this removal of a source of variability, the cVPT has been found to have relatively good reliability (Gladwin, 2018; Gladwin, Möbius, Mcloughlin, & Tyndall, 2018). A theoretical question is what is causing the bias. The task design implies that the cues serve as predictive stimuli for possible outcomes, or as a kind of prime (Kristjánsson & Ásgeirsson, 2019), due to some form of learning process (Failing & Theeuwes, 2018). However, it is conceivable that the visual features of the cues themselves acquire salience, as opposed to the theoretically motivating idea that predictive mechanisms determining the outcome of attentional shifting would result in the bias (Gladwin & Figner, 2014; Gladwin, Figner, Crone, & Wiers, 2011). Using a cVPT as an ABM task could provide evidence to help address this issue: if using a cVPT to train participants’ attention towards or away from an outcome indeed results in a bias involving the stimulus categories, rather than the specific cues used during training, this would suggest that the cVPT involves outcome-related processes rather than cue-specific learning.

Second, positive effects have been reported of what would usually be considered control conditions of ABM, in which no specific bias was induced but probes had a random relationship with emotional cues (Badura-Brack et al., 2015; Gladwin, 2017; Khanna et al., 2015). It has been suggested that whether a training variant makes emotional cues relevant or irrelevant to the training task may be an important factor in ABM (Gladwin, 2017). In usual sham conditions, emotional cues are irrelevant to the task and thus participants could be learning to ignore such stimuli when confronted by them. For instance, in a control condition of a training based on the Dot-Probe task, the location of emotional cues is non-predictive of the location of probe stimuli. In active training conditions, while the aim is to affect the direction of attentional biases, it is also usually the case that emotional information is relevant. In the Dot-Probe example, if probe locations are contingent on the location of emotional cues, then that makes those cues relevant. This could induce a “salience side-effect” in some designs: Participants may be learning to pay attention to the location of task-relevant emotional stimuli, even if the direction of attentional shifting is away from them. This could add noise and complexity to results, with different processes being affected in uncontrolled ways during training. In line with the idea that salience is an important factor in training, Approach-Avoidance Retraining for alcohol addiction reduced amygdala reactivity to alcohol stimuli (Wiers et al., 2015), which was interpreted as a neural signature of salience reduction.

The goal of the current study was to explore a novel form of ABM hypothesized to avoid this salience side-effect, which simultaneously may help understand the nature of the anticipatory spatial attentional bias. A training version of the cued Visual Probe Task was used, in which the probability of the location of probes relative to the outcome of cues is manipulated. This was termed predictive Attentional Bias Modification (predABM). To test whether this kind of predictive-cue training would affect attentional bias towards or away from actually presented emotional stimuli, a Towards Threat training condition and an Away from Threat training condition were compared using a normal Dot-Probe task post-training. As the delay between emotional cues and probe stimuli in Dot-Probe tasks is known to be potentially time-dependent (Mogg & Bradley, 2006; Noël et al., 2006), multiple cue-stimulus intervals were used. Note that due to the experimentally controlled random allocation of participants to groups, this post-only design allows valid statistical inference to be done: Statistically significant differences between groups on the post-test measures can be interpreted as an effect of training, with only the usual possibility of a false positive (which would also be present when analyzing difference scores). Beyond this basic point on the validity of randomized post-only designs, there are advantages and disadvantages to using a post-only versus pre-post design discussed further in the Discussion section. We hypothesized that training to shift attention towards versus away from the location of predicted upcoming threatening facial stimuli would affect attentional bias towards or away from such stimuli on the post-test stimulus-evoked attentional bias.

Method

Results

Table 1 shows descriptive statistics for the questionnaire data. Fifty-four participants were assigned to the Away from Threat group and 48 to the Towards Threat group.

Training effects on the Dot-Probe Task

Descriptive statistics for the Dot-Probe Task are shown in Table 3.

Table 3

Performance Data on the Dot-Probe task

Probe Location	CSI	Away From Threat	Towards Threat
Reaction time [ms]
Neutral	200 ms	505 (56)	502 (55)
Neutral	1,200 ms	516 (53)	511 (50)
Angry	200 ms	504 (58)	497 (53)
Angry	1,200 ms	523 (56)	511 (52)
Accuracy
Neutral	200 ms	0.96 (0.051)	0.95 (0.047)
Neutral	1,200 ms	0.97 (0.052)	0.96 (0.042)
Angry	200 ms	0.96 (0.048)	0.97 (0.045)
Angry	1,200 ms	0.97 (0.046)	0.97 (0.04)

Note. Means and standard deviations for reaction time and accuracy on the Dot-Probe task. Measure refers to the performance measures reaction time and accuracy. Probe location refers to the location where the probe stimulus appeared: The location of the cue where Neutral faces versus the cue where Angry faces would appear on non-probe trials. CSI refers to Cue-Stimulus Interval, the delay between cue presentation and probe presentation. Away from Threat and Toward Threat refer to the training conditions.

On RT, the hypothesized effect was found of Group × Probe Location, F(1, 91) = 4.75, p = .033, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = .05, shown in Figure 1.

Click to enlarge

Figure 1

Post-training RT bias per training group.

Note. The Figure shows the attentional bias, RT for Angry minus RT for Neutral, following the Towards Threat and Away from Threat training. The groups showed a relative shift in bias as expected.

The Towards Threat group had a bias towards threat relative to the Away from Threat group. The direction of the effect of Probe Location was reversed as expected between the groups, with shorter RTs on the Neutral than on the Threat location in the Away from Threat group, and shorter RTs on the Threat than on the Neutral location in the Towards Threat group. We do note that the magnitudes of the biases were small however, and the main effect of Probe Location did not reach significance in either group separately, despite the significant Group × Probe Location interaction. Further, an effect of CSI was found, F(1, 91) = 7.95, p = .0060, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = .08, due to slower responses at the long versus short CSI. No effects on accuracy were found.

Discussion

The aim of the current study was to provide a first test of the effects of predABM, a novel version of ABM using predictive cues. Rather than being trained to direct attention towards or away from threatening stimuli, participants were trained to direct attention towards or away from locations based on cues predicting where a threatening stimulus could appear. Thus, the training did not involve a direct stimulus-response association between stimuli in the threat category and attentional shifting, a feature of usual training tasks that could result in unexpected effects involving salience. The question was whether training using predictive cues would be able to affect stimulus-evoked attentional bias.

Performance data during training blocks showed that participants responded to the outcome-based task contingencies as expected. Responses were faster to probes appearing at the trained location. Note that this could reflect either an association involving the specific predictive cues or an association involving the stimulus category predicted by the cues—initial cVPT studies (Gladwin, 2016; Gladwin et al., 2018; Gladwin & Vink, 2018) were not able to distinguish between such possibilities concerning underlying mechanisms. Whether the latter, outcome-focused kind of association occurred was tested by the post-training generalization to the Dot-Probe task described below. Training effects were in fact found on the Dot-Probe task presented after the training. The Towards Threat group and Away from Threat group showed the expected relative decrease and increase, respectively, in reaction time for probes on Angry versus Neutral locations. Thus, the attentional response to emotional stimuli was changed via the stimulus categories of the outcomes of the predictive cues during training. Essentially, therefore, it was not the case that participants only learned to shift attention towards or away from the specific predictive cues. The results show that training involved the stimulus categories that were predicted by the initially visually neutral cues, even though the emotional stimuli never appeared on the same trial as the probe stimuli.

The results thus provide first support for the potential use of predABM. Although concern for salience side-effects in ABM, due to the informativeness of emotional cues, is as yet a recent development, the predABM provides a method that appears to be able to address this potential problem. However, we note that the potential training value of using an anticipatory attentional bias based on upcoming emotional stimuli, rather than responses to already-presented emotional stimuli, does not only depend on the salience side-effect. Anticipatory or preparatory processes related to emotional stimuli could be an interesting target for training in themselves, as this may have different effects from ABM involving stimulus-evoked processes. Further, a feature of predictive cues is that a wide range of possible stimuli can be associated with single conditioned cue. An interesting direction for future research is whether this may improve generalization to other stimuli, since attention is directed towards an abstract category rather than a specific set of stimuli.

A limitation of the current study is that only a single session was used, while effects of multiple sessions are likely most relevant for potential clinical applications and could provide larger effect sizes. However, the current results provide a proof-of-principle that the outcome-focused cued training task was able to change attentional processes related to the predicted stimulus category. A further limitation is that the population involved a sample of students. Patient groups are clearly an important target population, and it remains to be determined whether non-student samples respond to the training contingencies in the same way. A concern with training methods, especially for future use in clinical populations, is their impact on patients. The current study was also limited in its use of computer-generated angry faces as emotional stimuli: It cannot be assumed that the effects will generalize to other stimulus categories. Different results might be obtained in future research with, for example, stimuli representing physical threat, or verbal stimuli designed to evoke shame or guilt. Concerning the design, only a post-training assessment task for attentional bias was used, similarly to analyses involving post-training effects in previous studies (e.g., Gladwin et al., 2015). We note that, while pre-post designs have the advantage of providing a pre-training measurement, the logic of a post-training experimental design, with random assignment, is equally valid statistically: The chance of the groups having training-independent differences in attentional bias at post-test at random is the same as the chance of groups having training-independent changes in attentional bias from pre- to post-test at random. Further, a post-test design avoids test-retest effects, which could be a source of noise. There may also be theoretical reasons to expect effects to be caused on post-test states, rather than on pre-post shifts. Thus, while arguments can be made for either design, there is no reason to consider the lack of a pre-test a particular threat to the validity of conclusions drawn from the results. Finally, more work is needed to further explore the nature of training effects. Psychophysiology or neuroimaging methods could help test hypotheses on which underlying processes are affected, such as cue reactivity measures indicating changes in salience (Wiers et al., 2015) or attentional control (Eldar & Bar-Haim, 2010).

In conclusion, training to shift attention based on the expected stimulus-locations induces changes in attentional biases. The use of predictive cues in training may open interesting directions for further study.