Anxiety and attention to threatening pictures

Transcription

1 THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2001, 54A (3), Anxiety and attention to threatening pictures Jenny Yiend and Andrew Mathews MRC Cognition and Brain Sciences Unit, Cambridge, UK Previous research using attentional search tasks has revealed an anxiety-related bias favouring attention to threatening words when they are presented simultaneously with emotionally neutral words. In Experiment 1, using a similar task, a related effect was found here with emotionally threatening pictures. When pictures were used as location cues in a second experiment, high-trait anxious individuals were slower than less anxious controls when responding to targets requiring attentional disengagement from threat, and they were slower in general with pictures judged to be highly threatening. In a third experiment using the same task but with a longer cue exposure, a related disengagement difficulty occurred across both groups, although the more general slowing with severe threat was again confined to the anxious group. We conclude that attentional bias involves both a specific difficulty in disengaging attention from the location of any threat and a more general interference effect that is related to threat level. Previous research has suggested that threatening words (e.g., disease, failure) are more likely to capture attention than are emotionally neutral ones, particularly in anxiety-prone individuals (Mathews & MacLeod, 1994). This conclusion is based on two main sources of evidence. First, in interference tasks such as the emotional Stroop, words with a threatening meaning often cause slowed colour naming relative to matched neutral words (Williams, Mathews, & Macleod, 1996). Second, in attentional search tasks, targets may be detected faster when they occur in the same location as threatening words than in the location of emotionally neutral words (Broadbent & Broadbent, 1988; MacLeod, Mathews, & Tata, 1986). In the latter studies, both threatening and neutral words were displayed together for around 500 ms and were then followed in some trials by a probe stimulus (a small dot) to which subjects responded with a button press. The difference in latency to detect the probe in the prior location of a threatening versus a neutral word was taken as an index of which word was more likely to receive attention. Preferential attentional capture by threat words has been reported both in clinically anxious groups (MacLeod et al., 1986) and in high-trait anxious non-clinical populations (Broadbent & Broadbent, 1988), particularly when under stress (MacLeod & Mathews, 1988; Mogg, Mathews, Bird, & MacGregor-Morris, 1990). Typically, Requests for reprints should be sent to Andrew Mathews or Jenny Yiend, MRC Cognition and Brain Sciences Unit, 15 Chaucer Road, Cambridge, CB2 2EF, UK. The work reported here forms part of a thesis submitted by the first author to the University of Cambridge in partial fulfilment of the requirements for a PhD degree. We thank Brendan Bradley and Karin Mogg for their help and guidance in the early stages of this research. Ó 2001 The Experimental Psychology Society DOI: /

2 666 YIEND AND MATHEWS low-trait anxious individuals do not show the same response to threatening words and, indeed, have sometimes tended to show the reverse effect, suggestive of avoidance (MacLeod & Mathews, 1988; Mogg, Bradley, & Hallowell, 1994). These results have been used to argue that individual variations in selective attention to threatening cues when under stress may underlie vulnerability to anxiety states, by enhancing the acquisition of information about potential dangers in some people and minimizing it in others (Mathews & MacLeod, 1994). This argument implies that the findings with threatening words can be extended to other types of stimuli or events, leading to selective intake of information about danger by anxiety-prone individuals in everyday life, and thus representing a causal factor in anxiety disorders. If this were the case, we would expect to find that high-trait anxious individuals show attentional priority towards more realistic threatening stimuli, such as pictures or real events. We assume here that preferential processing of threat is a normal adaptive mechanism, which has evolved because it is important for detecting and avoiding dangers. However, we also assume that it is not so adaptive if any cues related to threat, no matter how weakly, always capture attention in preference to other information. If weak cues that require no immediate action, such as emotional words, are often successful in capturing attentional resources, then the ability to perform other tasks without interruption will be compromised. We have therefore suggested that some decision threshold may exist, such that below this threshold, stimuli related to threat are ignored (and perhaps inhibited), so that attention is likely to be captured only when the threshold is exceeded (Mathews & Mackintosh, 1998). Attentional differences associated with anxiety can thus be understood as arising from variations in this threshold level. If anxiety-prone individuals have a relatively low threshold, then their attention will be captured by weak threat cues. In contrast, non-anxious individuals are less likely to attend to weak cues and may even appear to inhibit them. Cues that are strongly related to threat, however, should capture attention in everyone, irrespective of anxiety level. A very similar hypothesis was put forward by Mogg and Bradley (1998), together with some preliminary supporting evidence (p. 836). Consequently, one question that we set out to investigate in the present experiments was that of whether stimuli rated as highly threatening would capture attention universally, whereas those rated as only mildly threatening would do so only with anxiety-prone individuals. One problem with addressing this question using word stimuli is that words may not vary sufficiently in affective intensity. Although words can be unequivocally negative in valence, it is far from clear that they can constitute a severe or highly salient threat, sufficient to elicit the supposed vigilant response even from low-trait anxious individuals. The stimuli chosen for the current study were thus a set of pictures that varied widely in subject and valence, and that reliably elicited distinctive subjective and physiological reactions (Lang, Bradley, & Cuthbert, 1995). In summary, research to date provides insufficient evidence to determine how attention to threatening cues may be affected by changing the type or the threat value of the stimuli, in either anxiety-prone or in non-anxious individuals. In Experiment 1 we tested the hypothesis that threatening pictures will capture attention when they compete for processing priority with matched neutral pictures, in anxiety-prone individuals. Emotional pictures were divided into those receiving high versus moderate threat ratings, to investigate the possibility that the

3 ATTENTION TO THREATENING PICTURES 667 most threatening stimuli would capture attention universally, with less threatening stimuli capturing attention only in anxiety-prone individuals. Method Participants EXPERIMENT 1 Participants were drawn from the student population of the University of Cambridge and were between 18 and 35 years of age, with a mean age of 22 years. Volunteers were solicited by letter that included two screening questionnaires (the 20-item Pittsburgh revision of the Taylor Manifest Anxiety Scale, Bendig, 1956; and a 10-item short form of the Marlowe Crowne Social Desirability Scale, Strahan & Gerbasi, 1972). Of 196 responders, we selected those who had extreme anxiety scores together with low social desirability scores (high social desirability scores may indicate invalid responses on anxiety questionnaires). Despite this screening, on full testing (see later) one low-anxiety volunteer was found to have a high social desirability score (two standard deviations above the group mean) and was therefore dropped. A total of 40 participants completed the experiment, 21 with high anxiety scores (13 female and 8 male) and 19 with low anxiety scores (11 female and 8 male). By design, the high-anxious group had significantly higher trait anxiety scores than the low-anxious group on the trait scale of the State Trait Anxiety Inventory (STAI; Spielberger, Gorsuch, Lushene, Vagg, & Jacobs, 1983); 58.9 versus 33.8, t(38) = 11.67, p <.001. They also had significantly higher scores on the state anxiety scale (53.8 vs. 37) and on the Beck Depression Inventory (Beck, Ward, Mendelson, Mock, & Erbaugh, 1961: 15.6 vs. 5.3), but had lower social desirability scores (Crowne & Marlowe, 1960; 11 vs. 14.3). Design Two groups of volunteers, scoring either high or low on self-report measures of anxiety, viewed threatening/non-threatening picture pairs; presented together on a computer screen for 500 ms. One of two target stimuli then appeared in the location previously occupied by one of the pictures, and participants responded by pressing a corresponding key. Latencies to detect these targets were used to index the extent to which the groups selectively attended to either highly or mildly threatening pictures. Materials and apparatus The stimulus set consisted of 144 digitized, monochrome pictures selected from the International Affective Picture System (IAPS; Lang et al., 1995). Of these, 48 were classified as threatening (24 highly threatening and 24 only mildly so) and 96 as non-threatening, on the basis of the judgements of six independent students raters (four male, two female). Each picture was judged on three scales according to how threatening (defined as emotionally disturbing or as making the viewer feel nervous), how pleasant, and how clear each one was, all using scales ranging from 0 (not at all) to 8 (extremely). Typical threatening pictures were photographs of corpses, weapons, assaults, or dangerous animals, and non-threatening pictures were typically domestic scenes or landscapes. Each threatening picture was matched to a non-threatening picture of similar orientation, brightness, and level of detail, as judged by eye. Content was taken into account so that the presence of people, animals, or inanimate objects was matched. A total of 24 neutral filler pairs were also prepared, resulting in a total of 72 pairs, of which 48 were threat/non-threat pairs: 24 with a picture judged as highly threatening

4 668 YIEND AND MATHEWS and 24 as only mildly so. Each of these latter subsets of 24 pairs was balanced for position (left or right) of the critical threat picture. In addition, 20 practice trials and 3 buffer trials were prepared using stimuli that had not met the rating criteria for inclusion in the main set. When presented, the visual angle between the centres of pictures in a pair was 6.5. Target stimuli were pairs of dots, using either a colon or two horizontally adjacent full stop keyboard characters. Each type of target was balanced for left right presentation position within each stimulus set. They were positioned approximately 3.2 from fixation and during the experiment were thus approximately in the centre of one of the previously displayed pictures. The task was presented on a 10.5² 8² colour monitor, using a 486 PC compatible computer. Presentation was controlled using MEL Version 1.0 software (Micro Experimental Laboratory; Schneider, 1988), and responses were collected by a MEL Version 1.0 parallel response box. Procedure In order to minimize eye movements, instructions given to participants emphasized that they should keep their eyes fixed on the central cross at all times, and these instructions were repeated at each break in the task. Additionally, they were instructed to respond to target stimuli as quickly as possible by pressing buttons corresponding to the orientation of the target. A total of 20 initial practice trials were followed by 3 buffer trials and then 72 experimental trials presented in a fully randomized order generated automatically for each subject by the MEL programme. Individual trials consisted of a fixation cross, presented for 1 s, followed by a picture pair displayed for 500 ms. Targets were then presented on the left or on the right and remained on screen until the response. Inter-trial intervals varied randomly between 500, 750, 1,000, or 1,500 ms. Finally, participants completed the full version of the State Trait Anxiety Inventory, the Beck Depression Inventory, and the Social Desirability Questionnaire. Results Judges ratings of picture sets Table 1 shows judges ratings for pictures designated as highly or mildly threatening, with the former being rated as both more threatening and less pleasant than the latter, but not differing in clarity. Both sets of threat pictures were rated as more threatening and less positive than their respective non-threat matching pictures (p <.001), but were not significantly TABLE 1 Mean threat, valence, and clarity ratings for threatening (critical) and matched (neutral) pictures used in Experiments 1 3 Severe set Mild set Rating M SD M SD Significance Critical pictures threat <.001 valence <.001 clarity ns Matched pictures threat ns valence ns clarity ns

5 different in clarity (p >.06, two-tail). The matching (non-threatening) sets did not differ significantly from each other on any dimension. Attentional task Trials with errors totalled 3.5% of the critical data (trials involving threatening pictures) and were excluded. A further 0.7% of responses with latencies greater than 1,100 ms were removed as outliers, on the basis of a box plot of the distribution. A mixed-design analysis of variance (ANOVA) was then carried out on the means of the remaining values, with anxiety group as the between-subjects variable and threat value, threat position, and target position as within-subject variables. There was a main effect of threat value, F(1, 38) = 4.1, p <.05, due to slower responses on trials involving a highly rather than a mildly threatening picture (535 vs. 526 ms). No interactions involving threat value approached significance (all Fs < 2). However, there was a significant interaction of group by threat position by target position, F(1, 38) = 5.00, p <.05. This three-way interaction was decomposed by conducting separate analyses for each group. For the low-anxious group the threat position by target position interaction remained significant, F(1, 18) = 5.67, p <.05. Although Tukey tests did not locate specific differences between pairs of means, inspection of Table 2 suggests that the interaction reflects slower detection latencies in threat than in non-threat locations (542 vs. 529 ms, see Table 2). For the high-trait anxious group the effect was in the opposite direction (519 vs. 530) suggesting faster responses when targets replaced threatening relative to non-threatening pictures, but the interaction was far from significant, F(1, 20) = Discussion ATTENTION TO THREATENING PICTURES 669 Low-anxious participants were slower to respond to targets replacing threatening rather than non-threatening pictures, irrespective of their rated severity, consistent with attentional avoidance of threat stimuli. Although the high-anxious participants latencies were in the reverse direction, suggesting the expected attentional capture effect, this effect was not significant. TABLE 2 Mean probe detection latencies a from Experiment 1 Latency Threat Probe Anxiety group location location M SD High anxious left left right right left right Low anxious left left right right left right a In ms.

6 670 YIEND AND MATHEWS These results confirm the expectation that individual differences in anxiety level would influence attention to pictures, with the significant interaction being similar in form to that found with emotional words (Mathews & MacLeod, 1994). In previous experiments, clinically anxious groups, or high-trait anxious students under examination stress, were found to be significantly speeded in detecting targets in the location of threat words, whereas non-anxious controls have generally shown non-significant trends in the reverse direction. Here, the non-anxious group was selected as having particularly low anxiety scores, perhaps accounting for a stronger tendency to avoid threat. Similarly, the fact that our anxious group was not drawn from a clinical population, nor was under any special stress, may account for the fact that the tendency for them to attend to threatening pictures did not reach statistical significance. None the less, the form of the interaction is consistent with the conclusion that highanxious individuals are more likely to have their attention captured by threatening stimuli than are non-anxious individuals. It therefore seems likely that the previously documented tendency of highly anxious individuals to attend more to threatening information (or to avoid it less) than do less-anxious individuals, applies to pictures as much as to words. 1 It was notable, however, that the expected differences in responses to highly versus mildly threatening pictures did not occur, suggesting that spatial attention to threat is not a direct function of subjectively rated threat value. Perhaps attention is preferentially held at the spatial location in which a possible threat is detected, regardless of the level of threat that is eventually assigned to the stimulus once it has been fully identified. This implies that such attention localization effects are initiated at an early stage, prior to full stimulus identification, consistent with earlier conclusions (Mathews & MacLeod, 1994). Interestingly, however, there was a main effect of rated severity in the present experiment, although this did not interact with the location of either the threatening picture or the target. Thus, although the spatial localization of attention did not seem sensitive to subjectively rated threat value, increased interference provided evidence that higher levels of subjective threat were indeed encoded. It is possible that two different attentional processes are involved and are revealed in different ways. The subjective severity of a stimulus had a non-spatial effect on attention, perhaps by taking up cognitive resources in proportion to the level of threat perceived. Spatial attention effects, on the other hand, may be influenced only by the relative attentional priority given to each location. Thus a picture having features related to those of previously encountered threats could lead to attention (or avoidance) at the relevant location in a relatively all-or-none manner, prior to full identification, and independent of the threat value or meaning that is encoded after full analysis. Such a process could arise via the fast but inaccurate route for detecting previously encountered threat cues, directly from thalamus to amygdala and bypassing the cortex, that has been described by LeDoux (1995). In Experiment 2 we again investigated the effect of threatening pictures on spatial attention, but used a method that allowed the assessment of both the engagement and disengagement components of attention (Posner, 1988; Stormark, Nordby, & Hugdahl, 1995). The method employed in Experiment 1, which was intended to allow direct comparison with 1 Although we frame our discussion in terms of threat, an anonymous reviewer correctly pointed out that the threatening pictures may also have been more novel. In pilot work with pictures selected as being novel but nonthreatening, we found no attentional differences between anxiety groups, suggesting that it is threat rather than novelty that is responsible for the effects reported here.

7 ATTENTION TO THREATENING PICTURES 671 earlier work on word stimuli, could not reveal whether threatening pictures resulted in differential attentional engagement at relevant locations, or differential difficulty in disengagement when targets appeared at a different location. Either effect alone, or both acting together, could account for differential latencies to detect targets at the location of threatening versus non-threatening stimuli. EXPERIMENT 2 As indicated earlier, little can be inferred about the exact nature of the attentional processing underlying faster (or slower) target detection in the prior location of threatening versus neutral words. Visual attention can be decomposed into several component processes, including those of engagement and disengagement (Posner, 1988; Posner, Inhoff, Friedrich, & Cohen, 1987). In the paradigm typically used to assess these components, a cue (such as brightening a rectangle) occurs in one of two locations, followed by a target, which is in either the cued or the alternative location. Relative to a no-cue baseline, targets are detected faster when in the cued (or valid) location, and slower when in the non-cued (or invalid) location. Speeding on valid trials has been attributed to the benefits of attentional engagement with the cued location, and slowing on invalid trials to the costs arising from having to disengage attention from the cued location, move, and engage the other location. In an attempt to use this method with emotional stimuli, Stormark et al. (1995) presented threatening or neutral words as cues and reported that the emotional words increased the benefits to validly cued targets and the costs of invalidly cued targets, relative to neutral words. Stormark et al. did not employ a no-cue baseline condition in this study, and no information was obtained about individual differences. It is thus not entirely clear which of the reported effects would differ from a baseline condition, nor whether they occur universally, or only in those with elevated levels of anxiety. Taken at face value, however, these results suggest that selective attentional capture by threat cues may be due both to greater engagement and to greater difficulty in disengaging from them. We adapted this experimental method by providing a picture as a cue to the probable location of the target. Thus, in trials when the picture validly predicted target location, comparison of target detection latencies for threatening versus non-threatening picture cues can be used to assess any differences in attentional engagement due to type of picture. Conversely, in invalid trials, when the target appears in a different location from the picture, any differences in detection latency due to type of picture can be taken to reflect ease of disengagement. Method Participants Participants were recruited and selected as before, and 39 completed the experiment; 20 in the highanxious group, 17 female and 3 male; and 19 in the low-anxious group, 11 female and 8 male. Two male members were excluded as before on the basis of high social desirability scores. Participants were all between 18 and 34 years of age, with group means of 20 and 21 years, respectively. They were divided into high- and low-anxious groups on the basis of their STAI-trait scores at the time of testing. Thus the high-anxious group had significantly higher trait anxiety scores than did the

8 672 YIEND AND MATHEWS low-anxious group, 53.7 versus 35.7, t(37) = 9.3, p <.001. As before, they also had higher state anxiety, 44.2 versus 33.2, and BDI scores, 10.7 versus 5.7, but lower social desirability scores, 9.6 versus Design There were three types of trial: valid, invalid, and no-cue. On valid trials a target arrow, pointing either up or down, appeared in the same location as the preceding picture cue, to the left or right of a central fixation cross. On invalid trials the target arrow appeared in the spatial location opposite the preceding picture cue. On no-cue trials a picture was not presented, and the screen remained unchanged for an equivalent duration before the appearance of the target. Of a total of 240 trials, 180 were critical, of which 60 were valid, 60 were invalid, and 60 were no-cue trials. In addition there were 60 valid filler trials not included in the analyses, to maintain the predictive validity of the valid trials. For each type of critical cued trial, 30 pictures were threatening, and 30 were non-threatening. All trial types were counterbalanced for target type and (when appropriate) type of picture. Materials Critical picture stimuli were drawn from the same set as those used in Experiment 1, but each picture was presented individually as a cue. There were a total of 60 threatening and 60 non-threatening pictures used in the 180 critical cued trials, plus another 60 used in the filler picture trials. In addition, 12 threatening and 12 non-threatening practice picture cues were assembled, together with a further 2 (one of each valence) for the buffer trials. The display at the start of each trial showed a central fixation cross, flanked on the left and right side by two white rectangular borders showing the two possible locations where pictures could appear. This fixation display remained on screen throughout each trial. Targets were black open-headed arrows displayed on a white background square, pointing either up or down to correspond to the appropriate response key. The visual angle between picture centres (the target position) and the fixation cross was again 3.2. Equipment, software, and data collection were identical to those of Experiment 1. Procedure Instructions presented on the screen asked participants to fixate the central cross at all times and to respond to target arrows by pressing the corresponding button as quickly as possible according to whether the arrow was pointing up or down. They were also told that on trials involving pictures, the picture location indicated the most likely position of the following target. Individual trials started with the fixation display, which lasted for 1,000 ms. Then, in valid and invalid trials, a picture cue was superimposed on the display for 500 ms, centrally located within one of the two borders on either the left or the right of the fixation cross. In no-cue trials the display was left unchanged for a further 500 ms. In other trials, the target arrow was further superimposed on the display, centrally located in one of the two borders, and remained on screen until the response. The time between picture-onset and target-onset was therefore 500 ms, although the pictures remained on screen until the end of each trial. The inter-trial interval varied randomly between 500, 750, 1,000, or 1,500 ms. A total of 32 practice trials (16 valid, 8 invalid, and 8 with no cue) were followed by 3 buffer trials and finally the 240 experimental trials, with an optional rest break at the halfway point. All trials were presented in a fully randomized order generated automatically for each subject by the MEL programme, and the entire task duration was approximately 20 min. Finally, subjects completed the same set of questionnaires as in Experiment 1.

9 Results ATTENTION TO THREATENING PICTURES 673 Response errors constituted 2% of the critical trials and were excluded. Latencies greater than 750 ms or less than 100 ms, totalling another 2% of the data, were excluded as outlying responses on the basis of a box plot. Our main hypothesis concerned the comparison of threatening versus non-threatening pictures within valid and invalid trials. The mean target detection latencies from these trials were therefore entered into an ANOVA, with one between-subjects factor (group: high vs. low anxiety) and two within-subject factors (trial type: valid, invalid; and cue type: threat, nonthreat). There were main effects of group (anxious participants were slower), of cue type (responses were slowed by threat), and of trial type (invalid slower than valid). Cue type also interacted with trial type and with group in two-way interactions, but these were all subsumed under a higher order interaction of group by trial type by cue type, F(1, 37) = 4.90, p < To clarify the meaning of this interaction, a further ANOVA was performed for each subject group separately with trial type and cue type as within-subject variables. For the lowanxious group this revealed only a trend for latencies in valid trials to be faster than those in invalid trials, F(1, 18) = 3.1, p <.10. The interaction with threat or non-threat cue type did not approach significance, F(1, 18) = For the high-anxious group, there were main effects of cue type and trial type, qualified by a significant interaction between them, F(1, 19) = 15.39, p <.001. Tukey (HSD) tests applied to these means revealed that valid threat, valid non-threat, and invalid non-threat trials did not differ significantly from each other, whereas the invalid threat trials were significantly slower than all the other three. For non-threat picture cues, the latencies for invalid and valid trials were almost identical (458 vs. 456 ms), whereas the equivalent difference for threatening picture cues was 25 ms (486 vs. 461 ms, see Figure 1 and Table 3). Figure 1. Target detection latencies in valid/invalid trials involving threatening and non-threatening pictures for high- and low-anxiety groups in Experiment 2. 2 A further ANOVA that included an additional factor of side of the screen on which the cue was presented (left vs. right) revealed only a trend for this three-way interaction to be further qualified by side, F(1, 37) = 3.05, p <.10. The largest slowing on invalid trials tended to be high-anxious subjects following threat cues presented on the right side of the screen.

10 674 YIEND AND MATHEWS The threat severity factor was not fully nested in the design, there being no equivalent to the high versus mild threat variation in the non-threat or the no-cue trials, so that some relevant comparisons were not included the analysis. For this reason, we assessed the effects of cued and no-cue trials in the absence of threatening pictures, by contrasting the no-cue trials with non-threat valid and non-threat invalid trials in a subsidiary analysis. There were no significant differences due to trial type (means were: valid 437, no-cue 442, and invalid 441 ms). We also analysed data from trials involving threatening pictures only, divided according to their rated threat value (high or mild). This revealed a main effect of severity, F(1, 37) = 4.73, qualified by an interaction with anxiety group, F(1, 37) = 7.25, p <.05. Tukey (HSD) tests showed that high-anxious participants were slowed more by high than by mild threat (483 vs. 464 ms), but that low-anxious participants were not (418 vs. 420 ms). The interaction with trial type (valid vs. invalid) did not approach significance, F < 1. Thus, as in Experiment 1, there was an interference effect associated with rated threat severity, which was not dependent on location. Unlike the previous experiment, however, this effect was present only in the highanxious group. Discussion TABLE 3 Mean probe detection latencies a from Experiment 2 Latency Anxiety group Trial type Cue type M SD High anxious valid threat non-threat invalid threat non-threat Low anxious valid threat non-threat invalid threat non-threat a In ms. The main purpose of Experiment 2 was to examine whether attentional engagement and disengagement, or both, would be affected by threatening picture cues. Although the data for neutral picture cues alone failed to demonstrate the expected pattern of costs and benefits relative to a no-cue condition, an analysis combining data from both neutral and threat cues did show slowing on invalid relative to valid trials. In low-anxious participants this effect was not qualified by type of picture. These participants were slower to detect targets in invalid trials, presumably due to the cost of having to shift attention to the non-cued location, and/or the benefit conferred by engagement with the cue. In high-anxious participants, however, the relative slowing on invalid trials was confined to those involving threatening pictures. Emotionality of the picture had little effect on latencies in valid trials (461 ms for threat vs. 456 for neutral), and these latencies were also no different from no-cue trials in this group (460 ms). On the basis of these data, there is no evidence to suggest that greater engagement with threatening pictures led to speeding. The significant

11 ATTENTION TO THREATENING PICTURES 675 slowing seen following invalid threatening cues (486 ms) thus indicates that the main effect of threatening pictures was to delay attentional disengagement. Attentional biases associated with anxiety could thus be due, at least in part, to associated differences in the ease of disengagement from threat. A possible complicating factor in arriving at unequivocal conclusions based on these data is that threatening pictures were also associated with a general slowing effect in highly anxious participants. We suspect that highly threatening pictures initiated task-irrelevant processing in these participants, related to emotional content or meaning, which led to the delays in task performance. The location-based effects may therefore have been confounded with this more general interference process, complicating interpretation of the data. Consequently, we cannot discount the possibility that a greater benefit would have been observed in the highanxious group due to engagement with the threatening pictures, had this not been masked by an interference effect in the opposite direction. Similarly, the costs due to invalid threat cues may have appeared to be larger than they would in the absence of interference. It seems unlikely that this can be an adequate explanation of the data, however, because interference due to threat severity was not at all influenced by target location. As in Experiment 1, it seems more likely that there are two distinct processes involved: an interference effect associated with threat severity but independent of location; and a separate location-dependent effect, associated with spatial allocation of attention but less influenced by severity. The results of Experiment 1 suggested that highly anxious individuals are less likely to avoid attending to the location of threatening stimuli than their low-anxious counterparts. Experiment 2 similarly indicated that highly anxious individuals have more difficulty disengaging attention from a threat location in order to engage elsewhere, even when this is required by the task. It seems plausible that this conclusion is limited to fairly early processing stages, however, for several reasons. First, as the threatening pictures are perceived as aversive, viewers particularly if they are anxious are presumably motivated to avoid giving them prolonged attention. Second, after attention has been disengaged from a cued location, the usual pattern of faster responses to validly cued targets is typically reversed, in favour of targets appearing in other locations. This reversal is taken to imply a phenomenon known as inhibition of return, in which a disengaged location becomes inhibited relative to alternative locations (Posner & Cohen, 1984; Riggio, Bello, & Umiltà, 1998). We investigated this question in a third experiment, identical in all respects to Experiment 2, except for the stimulus-onset asynchrony (SOA) between the cue and target. In Experiment 3 this SOA was increased from 500 to 2,000 ms, thus allowing ample time for attention to shift away from the picture cue. We expected that under these conditions invalid trials would be associated with faster probe detection times than would valid trials. Our main interest concerned the extent to which this would apply equally to the threatening and neutral picture cues, and to high- as well as low-anxious participants. It might be, for example, that full attentional disengagement from threatening pictures would prove difficult, particularly for highly anxious individuals, leading to an absence of inhibition of return. An alternative hypothesis is that initial attentional capture by threatening cues may be succeeded by greater avoidance in highly anxious individuals. In this view, once a threatening stimulus has been identified, the aversive reaction experienced by anxious individuals will motivate greater efforts to keep attention away from it. If this effort were to be successful, then inhibition of return may be correspondingly greater rather than being absent.

12 676 YIEND AND MATHEWS Method Participants and materials EXPERIMENT 3 A total of 40 participants were recruited as before, 16 male and 24 female, with gender balanced across anxiety groups. Ages were between 18 and 31 years, with the mean for the groups being 22.7 (high anxious) and 20.7 years. The two groups differed significantly on all questionnaire measures, with the high-anxious group scoring significantly higher than the low-anxious group on STAI trait anxiety (54.7 vs. 30), STAI state anxiety (39.3 vs. 30.6), and the BDI (14.6 vs. 4.6), but significantly lower on social desirability (10.1 vs. 13.3). The pictures used were identical to those in Experiment 2 and were presented using an updated version of the same software (MEL Version 2.0 including serial response box, Model 2A). Procedure The procedure was identical in most respects to that of Experiment 2. The single methodological difference was the lengthening of the cue-onset target-onset duration to 2,000 ms. The precise duration on any trial was chosen randomly from 1,950, 2,000, and 2,050 ms to prevent prediction of target onset. Results There were no latencies below 200 ms, so none were excluded as anticipatory responses. The cut-off for long outlying latencies was set at 750 ms, on the basis of a box plot of the distribution. In total 3.1% of the critical data were excluded: 1.3% due to slow responses and 1.8% due to subject errors. As before, the analysis of major theoretical interest concerned target detection latencies in valid and invalid trials, cued by threatening versus non-threatening pictures. A mixed-model ANOVA was thus carried out on mean latencies, with group (high or low anxiety) as the between-subjects variable, and trial type (valid, invalid) and cue type (threat, non-threat) as within-subject variables. This analysis revealed significant main effects of anxiety group and trial type (but now with invalid trials faster than valid trials; 428 vs. 435 ms), and a significant interaction involving trial type and cue type, F(1, 38) = 5.4, p <.05. The three-way interaction involving anxiety group did not approach significance. 3 Tukey (HSD) tests were used to compare means in order to clarify the interpretation of the interaction of trial type by cue type. Valid trials did not differ according to cue type (threat 434 ms, non-threat 436 ms), nor did they differ from invalid threat trials (432 ms), but they were significantly slower than invalid non-threat trials (424 ms). The interaction thus reflects the fact that trials with invalid non-threatening cues led to the predicted speeding, but trials with invalid threatening cues did not (see Figure 2 and Table 4). As in the analysis of data from Experiment 2, we also contrasted no-cue trials with both valid and invalid trials involving non-threatening pictures. This ANOVA revealed a main slowing effect of anxiety but, more importantly, a main effect of trial type, F(2, 76) = 13.11, 3 As before, a further ANOVA included the factor of which side of the screen the cue was presented (left vs. right). This revealed an interaction of group by side, F(1, 38) = 5.22, p <.03, but as this did not involve the type of cue it was not considered further.

13 ATTENTION TO THREATENING PICTURES 677 Figure 2. Target detection latencies in valid/invalid trials involving threatening and non-threatening pictures in Experiment 3. TABLE 4 Mean Probe detection latencies a from Experiment 3 Latency Anxiety group Trial type Cue type M SD High anxious valid threat non-threat invalid threat non-threat Low anxious valid threat non-threat invalid threat non-threat a In ms. p <.001. Tukey (HSD) tests revealed that responses to targets in valid trials were not reliably faster than those in no-cue trials (mean for valid 436 ms, no-cue 441 ms), but that both were significantly slower than responses in invalid trials (424 ms). Thus, in line with predictions based on inhibition of return, 2 s after the appearance of a neutral cue, targets were detected more rapidly at the alternative (non-cued) location. Also, as in Experiment 2, we analysed target detection latencies following threatening picture cues only, after dividing the latter into highly and mildly threatening sets. Other than the main effect of anxiety group, the only significant finding was of a group by severity interaction, F(1, 38) = 6, p <.05, which was not further qualified by trial type. Tukey (HSD) tests indicated no reliable difference between responses following severe or mild threat in low-anxious participants (416 vs. 418 ms). In high-anxious participants, responses after severe threat tended to be slower (454 vs. 442 ms), although just falling short of the 5% confidence level. The interaction can thus be interpreted as being due to a strong trend for the high-anxious group to be slower after severe threat, whereas the low-anxious group was not. This anxietyrelated interference, which was proportional to threat value but independent of target location, is essentially the same as that found in Experiment 2.

14 678 YIEND AND MATHEWS Finally, because the design of Experiments 2 and 3 was the same apart from the SOA involved, the main ANOVA was repeated on both data sets together, including an additional between-subjects factor of experiment (SOA of 500 vs. 2,000 ms). The three-way interaction of trial type by cue type by group was significant in the combined analysis, F(1, 66) = 5.2, p <.03, and there was a non-significant trend for this to interact with experiment, F(1, 66) = 4.25, p <.10. Discussion The first aim of Experiment 3 was to examine whether attentional disengagement would occur after longer cue exposures, leading to inhibition of return and corresponding faster target detection in the previously unattended location. With non-threatening picture cues this expectation was confirmed: Targets following invalid cues were detected reliably faster than targets that were validly cued or were not cued at all. The second aim was to investigate whether such consequences of disengagement would be less apparent with threatening picture cues, particularly in highly anxious participants. Results were consistent with an effect of threat on disengagement, but did not show clear differences due to anxiety. All participants were faster on invalid trials cued by non-threatening rather than by threatening pictures. It may be that disengagement from threatening pictures was less complete than that from non-threatening pictures, or that inhibition of return to the threatening location was less marked. Unlike the disengagement of attention from threatening pictures seen in Experiment 2, which was apparently difficult for high- but not for low-anxious participants, the related effect in Experiment 3 was similar in both anxiety groups. That is, after 2 s both groups showed an advantage when detecting targets following invalid non-threatening pictures, but not when detecting targets following invalid threatening pictures. This residual and persistent difference could reflect a universal difficulty in disengaging entirely from the location of a single threatening cue, regardless of anxiety level. It should be noted, however, that the interaction with anxiety group found in the combined analysis was not reliably qualified by experiment. This means that our conclusions about different effects with the longer SOA must remain tentative. There was a difference between anxiety groups, however, in the extent to which severity of threat slowed responding. As in Experiment 2, the more highly threatening pictures produced more interference than did milder pictures, but only in highly anxious participants. Also, as before, this interference effect was independent of target location and is thus difficult to attribute to the spatial distribution of attention. The severity-related interference effect can therefore be distinguished from the other main effect of threat found here, which did relate to probe location. GENERAL DISCUSSION The main finding across all three experiments was that of differential deployment of visual attention to threatening versus non-threatening pictures, that under some circumstances was associated with individual differences in anxiety level. Generally speaking, these anxietyrelated differences were similar in nature to those previously studied in experiments with

15 ATTENTION TO THREATENING PICTURES 679 emotional words. In Experiment 1, when two types of picture competed for attention, lowanxious individuals were relatively faster to detect targets in the location of non-threatening pictures. This was interpreted as suggesting that low-anxious individuals tend to avoid emotionally aversive stimuli in favour of alternatives, whereas highly anxious individuals do not. The results of Experiment 2 further suggest that at least part of this anxiety-related contrast is attributable to differences in early attentional disengagement from threatening and non-threatening stimuli. Anxiety-prone individuals were particularly slow to detect targets that followed invalid threatening picture cues. These participants apparently disengaged rapidly from non-threatening pictures, but failed to do so in the case of threatening pictures. Low-anxious individuals, in contrast, demonstrated much less distinction in target detection latencies as a function of different picture cues. Although this finding does not explain the attentional preference for non-threatening pictures shown by the non-anxious group under conditions of attentional competition in Experiment 1, it does offer a possible explanation for the lack of this effect in highly anxious individuals. In this latter group, we assume that attention to a threatening stimulus would not be readily disengaged, relative to a non-threatening stimulus. Differential disengagement rates could thus account for a tendency for highly anxious individuals to be relatively faster when detecting probes in the location of threatening stimuli. Experiment 3 suggested that this anxiety-related difference had largely dissipated after 2 s. Threatening pictures still retained some capacity to impede disengagement, but the difference between groups in this respect was no longer significant. This suggests that (in the absence of competition) threatening pictures tend to hold attention at early processing stages, with the strength of this effect varying according to anxiety level. The tendency of a single threatening stimulus to hold attention presumably opposes the normal process of redeployment of attention and inhibition of return. If so, then we could suppose that the attentional hold of threat was greater in highly anxious individuals at 500 ms, whereas by 2,000 ms this difference had been eroded and resembled that seen in the low-anxious group. One problem with this account is that it implies that valid threat trials should have been faster than valid non-threat trials, due to greater engagement at the valid location. Contrary to this prediction, there were no such threat/non-threat differences in either Experiment 2 or Experiment 3. However, it should be recalled that threat trials tended to be associated with generally slower latencies, particularly in anxious participants. If one were able to discount this interference effect, then latencies in valid threat trials might indeed appear faster in anxious individuals. Another potential limitation of our attentional account of these results is that we did not monitor eye movements. In the absence of such measures, we do not know whether participants moved their eyes in some trials, despite our instructions to the contrary. However, even if so, we do not believe that our general conclusions would have been very different, as eye movements are usually assumed to follow attention shifts, rather than vice versa. There was a striking absence in these data of any effects due to subjective threat severity, which was never associated with the location of attention. It seems that whether a threat is subjectively evaluated as highly or only mildly aversive has little influence over whether or not it holds attention at that location, at least in the first 2 s of processing. The most plausible explanation seems to be that all threatening stimuli, within the range sampled here, tend to hold attention at their location (or fail to do so) in an all-or-none manner. Evidence discussed elsewhere (e.g., Mathews & Mackintosh, 1998; Öhman & Soares, 1994) indicates that threatening

16 680 YIEND AND MATHEWS stimuli can be detected pre-attentively, causing interference, priming, or autonomic effects, even when their nature cannot be reported. We assume that such detection initiates processes leading to attentional localization, prior to any detailed analysis of the threat content or meaning. Such localization would presumably facilitate the conscious assessment of danger severity, rather than being a consequence of it. It follows that attentional localization precedes awareness of threat value, rather than the other way round. Nonetheless, our data show that participants, particularly those with high-trait anxiety levels, did encode threat value. Irrespective of target position, highly threatening pictures were associated with slower responses than were milder pictures in Experiment 1. In Experiments 2 and 3, similar interference was confined to the high-anxiety group. We suggest that this location-independent effect arose from task-irrelevant processing concerned with threat meaning and severity, taking up more cognitive resources as threat value increased. As well as being independent of attentional location, anxiety-related interference seemed persistent, as the groups still differed after 2 s. Group differences in attentional location and interference may thus follow somewhat different time courses. Attentional location differences between groups were significant only at 500 ms, despite persisting overall effects of picture type at 2 s. Interference effects that were independent of location were also present at 500 ms in both groups (Experiment 1) or only in the high-anxious group (Experiment 2), but persisted at 2 s only in the latter group (Experiment 3). Greater persistence of interference in high-anxious individuals is consistent with the suggestion that it reflects more extensive processing of threat meaning. Related to this suggestion, it seems likely that there was indeed a differential emotional reaction to highly versus mildly threatening pictures. There is good evidence that startle to a loud noise is enhanced when viewing unpleasant pictures taken from the set used here, and that startle augmentation is greater in anxious individuals, or when pictures are perceived as threatening to the viewer (e.g., Corr, Kumari, Wilson, Checkley, & Gray, 1997). Furthermore, it is known that fear-related modulation of startle is directly controlled by the amygdala (Davis, 1989). Thus, viewing highly threatening pictures is sufficient to elicit activation of brain systems intimately involved in fear reactions. Only rather limited evidence, however, has been found in support of our earlier hypothesis that anxiety is related to the level of threat severity sufficient to hold attention at that location. Threatening and non-threatening pictures do elicit different attentional responses, and these sometimes vary according to anxiety level. Rather than subjectively defined threat severity, however, our data suggest that anxiety is associated with attention being held at the location of any threatening stimulus. When only limited information about a threat is available, we suppose that highly anxious people are more likely to attend that location and may thus learn more about it. Anxiety-related response threshold differences based on sensitivity to threat severity alone seem inadequate to explain these findings. Rather, we suggest that the presence of a threat may lead anxiety-prone people to adopt a vigilant checking mode, even prior to full information being available. In non-anxious people, the same event is likely to be followed by indifference, or active avoidance. Any threshold difference between individuals is thus thought to relate to the readiness with which such a vigilant mode is adopted, rather than the threat value necessary to elicit it.