Personality and Individual Di€erences 28 (2000) 1063±1077 www.elsevier.com/locate/paid Trait anxiety, defensiveness and selective processing of threat: an investigation using two measures of attentional bias Karin Mogg*, Brendan P. Bradley 1, Claire Dixon, Susan Fisher, Helen Twelftree, Andrew McWilliams Department of Experimental Psychology, University of Cambridge, Downing Street, Cambridge, CB2 3EB, UK Received 14 December 1998; received in revised form 21 June 1999; accepted 26 June 1999 Abstract Attentional biases for threat were examined in a non-clinical sample (N = 60), with each participant tested on both the modi®ed Stroop colour-naming and dot probe tasks. Three groups were selected on the basis of trait anxiety and social desirability scale (SDS) scores: ``low anxiety'' (LA: low trait, low SDS), ``repressor'' (REP: low trait, high SDS) and ``high anxiety'' (HA: high trait, low SDS). Results from the colour-naming task suggested that high levels of defensiveness (in combination with low trait anxiety) were associated with greater avoidance of threat. The REP group showed less interference in colour-naming threat than neutral words; whereas the HA group showed increased interference due to threat words. On the dot probe task, there was a general tendency for this non-clinical sample as a whole to show avoidance of social threat relative to neutral words, but there was no bias for physical threat words. Avoidance of social threat was signi®cant only within the REP group. No relationships were found between the measures of cognitive bias from the two tasks, suggesting di€erent underlying mechanisms. Results are discussed in relation to previous ®ndings and theoretical views of the e€ects of anxiety and defensiveness on the processing of threat. 7 2000 Elsevier Science Ltd. All rights reserved. Keywords: Anxiety; Defensiveness; Attentional bias * Corresponding author. 1 Correspondence may also be addressed to Brendan P. Bradley 0191-8869/00/$ - see front matter 7 2000 Elsevier Science Ltd. All rights reserved. PII: S 0 1 9 1 - 8 8 6 9 ( 9 9 ) 0 0 1 5 7 - 9 1064 K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 1. Introduction There has been considerable interest in research into attentional biases for threat information in anxiety because recent cognitive theories have proposed that such biases may play a key role in the development and maintenance of clinical anxiety states (e.g., Mathews & MacLeod, 1994; Eysenck, 1992; Williams, Watts, MacLeod & Mathews, 1997). It is well-established that clinically anxious individuals show an attentional bias towards threat cues (e.g., reviews by Mathews & MacLeod, 1994; Mogg & Bradley, 1998). If such biases do indeed provoke or intensify clinical anxiety states, then it may prove useful for anti-anxiety treatments to target them; i.e., to encourage anxious patients to adopt a more ``normal'' attentional style. However, a potential problem with such an approach is that there is considerable uncertainty about what constitutes a normal pattern of bias, as the speci®c emotional and personality variables that determine the presence or absence of attentional biases for negative information in the normal population are not well understood. One key issue is whether normal, low anxious individuals have a tendency to divert their attention away from unpleasant information, or whether they have no attentional bias, either towards or away from threat. A tendency to disregard trivial unpleasant stimuli in the environment might serve at least two useful functions. First, it would seem helpful in maintaining attention on current goals and in protecting against the distracting e€ects of minor task-irrelevant negative cues. Second, it may assist in mood regulation; that is, in reducing anxious mood and in maintaining positive mood states. Experimental evidence of avoidance of mildly aversive information has been suggested by several studies (e.g., avoidance of threat words or faces in normal or non-dysphoric samples; MacLeod, Mathews & Tata, 1986; Bradley, Mogg, Millar, Bonham-Carter, Ferguson, Jenkins & Parr, 1997), although it does not seem to be a very robust e€ect given that it has not been universally found in studies of attentional bias. One reason why avoidance of threat may be dicult to demonstrate reliably is that it may not simply be a function of low anxiety, but it may also be determined by other personality variables, such as defensiveness. One proposal is that there are two subtypes of low trait anxious individuals, who can be discriminated on the basis of scores on the Social Desirability Scale (SDS; Crowne & Marlowe, 1964). High SDS scores are assumed to re¯ect high levels of defensiveness, a high need for social approval, and a reluctance to report negative emotional states. Thus, one may distinguish between low trait anxious individuals with low levels of defensiveness (i.e. ``true low anxious''), and those with high levels of defensiveness (i.e., socalled ``repressors''). Weinberger, Schwartz and Davidson (1979) found that repressors exhibited greater somatic responses to stress than did low trait anxious individuals, suggesting that repressors are physiologically anxious despite their self-reports of low subjective anxiety. Thus, repressors may resemble high trait anxious individuals in terms of their physiological and cognitive responses to threat, but may consistently under-report their anxiety levels because of concern to make a positive impression on others. An alternative view is that repressors have a defensive bias in selective attention; i.e., a cognitive ®lter, or schema, which results in avoidance of negative or threatening cues (Bonanno & Singer, 1990). Indeed, Eysenck (1997) speci®cally proposed in his new cognitive theory of anxiety that repressors have a cognitive bias opposed to the processing of threat, whereas high trait anxious individuals K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 1065 have a bias in favour of threat, and low trait anxious individuals have no bias (see also Derakshan & Eysenck, 1997). Experimental evidence of the relative e€ects of trait anxiety and defensiveness on attentional biases is mixed. Such research has typically compared three groups: low anxiety (LA; low anxiety, low SDS), repressor (REP; low anxiety, high SDS) and high anxiety (HA; high anxiety, low SDS). The fourth combination, so-called ``defensive high anxiety'' (i.e., high anxiety, high SDS) seems to be relatively uncommon (Weinberger et al., 1979). In one study, using a modi®ed Stroop task to investigate processing biases for threat, Dawkins and Furnham (1989) found that REP and HA were slower to name the colours of threat than neutral words, whereas LA showed no e€ect. These results suggested not only di€erent patterns of cognitive bias for threat within those reporting low trait anxiety (i.e., LA and REP), but that repressors resembled high trait anxious individuals in showing vigilance for threat. However, Fox's (1993) study suggested the opposite conclusion. She used a dot probe task to assess attentional bias, in which pairs of words were presented for 500 ms, one word above the other, and participants read aloud the upper word. On occasional trials, a dot probe appeared in the location of one of the words, and participants were required to respond to the probe as quickly as possible by pressing a key. The rationale for the task is that participants are faster to respond to probes in attended rather than unattended spatial locations (e.g., Posner, Snyder & Davidson, 1980); so the pattern of RTs to probes indicates the allocation of attention to the preceding words. This task had been previously used by MacLeod et al. (1986) to demonstrate an attentional bias for threat words in generalised anxiety disorder (GAD). Fox found that the HA group were faster to respond to probes replacing social threat words than neutral words, consistent with vigilance for threat. By contrast, the REP group was relatively slower to detect probes replacing social threat words, which suggested attentional avoidance of social threat. The LA group showed no bias. Further evidence that was consistent with an avoidant bias in repressors was suggested by a dichotic listening study where repressors showed less disruption by negative words on the unattended channel than high trait anxious individuals (Bonanno, Davis, Singer & Schwartz, 1991). One variable that might account for such discrepant ®ndings is the use of di€erent tasks to assess attentional biases. Evidence suggesting vigilance for threat in repressors was obtained from the card version of the Stroop, but this task has some drawbacks. Although interference e€ects are typically interpreted in terms of attention being preferentially allocated to the threat content, this interpretation is not entirely certain, because the interference e€ect may not necessarily occur at the input stage of information processing, but at a later response selection stage (e.g., C.M. MacLeod, 1991). Further interpretative diculties with the Stroop task have been noted by de Ruiter and Brosschot (1994). On the one hand, interference e€ects might re¯ect attentional vigilance, i.e., impairment in colour-naming is due to preferential allocation of attention to the threat word content, consistent with the threat interference e€ects found in high anxious individuals. On the other hand, interference e€ects might re¯ect cognitive avoidance, whereby the impairment in colour-naming is due to increased processing resources and e€ort being required to direct attention away from negative information; which may account for the interference e€ects found in repressors. Their criticisms highlight the importance of not relying on just one task to assess attentional biases. The main aim of the present study was to investigate the e€ects of anxiety and defensiveness 1066 K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 on attentional biases for threat information in a normal sample. A novel feature of the study was that each volunteer was tested on two measures of attentional bias, so that we could examine, within the same sample of participants, whether the use of di€erent tasks may explain the discrepant ®ndings noted above. One task was a computerised version of the emotional Stroop task, similar to that used by Mogg, Bradley, Williams and Mathews (1993) to investigate selective processing of threat information in GAD. In this task, words were presented individually on a background ¯ash of colour, and vocal RTs and errors were recorded separately for each word (which has an advantage over the card version of the Stroop where RT data and errors are confounded). The second measure of attentional bias was obtained from the dot probe task. One advantage of this task is that it does not rely on interference e€ects to measure attentional biases, and so vigilance for threat should be evident from relatively faster responses to probes in the location of threat words. The task used here was similar to that used by Mogg, Bradley and Williams (1995), in which on each trial, a word pair was brie¯y presented, immediately followed by a dot probe in the location of one of words, and participants indicated the position of the probe as quickly as possible by pressing one of two response keys. This forced-choice version of the task was based on that used by Posner, Snyder and Davidson (1980, Experiment 3). One prediction was that high trait anxiety should be associated with greater vigilance for threat words relative to neutral control words, following from cognitive formulations of anxiety±congruent biases in selective attention. This e€ect should be evident from slower RTs to threat than neutral words on the Stroop task, and faster RTs to probes replacing threat than neutral words on the dot probe task. A second prediction was derived directly from Fox's (1993) ®ndings and Eysenck's (1997) theory Ð that is, repressors should show attentional avoidance of threat words. Moreover, such avoidance should be most apparent for words related to social threat (Fox, 1993). This e€ect should be evident from faster RTs to threat than neutral words on the Stroop task, and faster RTs to probes replacing neutral than threat words on the dot probe task. 2. Method 2.1. Participants One hundred and seventy four students volunteered for the study and returned by post a screening questionnaire consisting of 20 items from the Taylor Manifest Anxiety Scale (TMAS) and 10 from the SDS. The short-form TMAS was used for screening because it provides a convenient estimate of trait anxiety, with True/False response format compatible with the SDS, and it correlates 0.93 with the full TMAS (Bendig, 1956). The short-form SDS (Form X1, Strahan & Gerbasi, 1972) has been found to correlate 0.96 with the full SDS (Fischer & Fick, 1993). To minimise the proportion of the sample with mid-range levels of anxiety and defensiveness, those with extreme TMAS and SDS scores were favoured in recruitment (except for those with high scores on both, i.e., ``defensive high anxious''). There were 60 in the ®nal sample. K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 1067 2.2. Materials The stimulus words were taken from previous research into anxiety-related attentional biases (e.g., Fox, 1993; Mogg et al., 1993, 1995). There were two types of threat words: 48 related to physical threat (e.g., cancer, mutilated) and 48 related to social threat (e.g., despised, pathetic). Another 48 categorised neutral words (household-related words, e.g., toaster, furnished) served as control stimuli. The three word sets were matched for length and frequency using Carroll, Davies and Richman's (1971) norms. For the dot probe task, each word was paired with a length and frequency matched uncategorised neutral word (e.g., downwind, mayor) to form 144 word pairs. For the Stroop task, an additional 48 uncategorised neutral words (length and frequency matched with the categorised neutral words) were used as ®llers and to address a subsidiary issue regarding the e€ect of word-categorisation on colour-naming interference. Each set of words was divided into two equivalent lists, matched for length and frequency, so that each participant received one list (i.e., half the words) in the dot probe task, and the other in the Stroop task. The allocation of the word lists to the two tasks was counterbalanced across participants. Thus, for each person, 72 word pairs were presented in the dot probe task (i.e., 24 social threat, 24 physical threat and 24 categorised neutral words, each paired with an uncategorised neutral word), and 96 words were presented in the Stroop task (24 social threat, 24 physical threat, 24 categorised neutral, 24 uncategorised neutral). The tasks were run on a IBM 286 PC with MEL version 1 software (Schneider, 1988) and MEL response box with built-in voice key. 2.3. Procedure Half the participants received the Stroop task before the dot probe task, and vice versa for the other half. Each was seated about 60 cm from the screen. In the modi®ed Stroop task, there were 15 practice and three bu€er trials and 96 test trials presented in a new random order for each person. The participant started each trial by pressing the keyboard space bar. A ®xation box (15  35 mm) then appeared for 500 ms in the centre of the screen. This was replaced by a word in white uppercase letters (approx. 5 mm high) on a background patch of colour that was either red, green or blue. To restrict the display of the colour (which ®lled one line across the screen), each side of the screen was masked with black card (11  3.5 cm). The colour background and word appeared simultaneously, with the colour displayed only for one screen refresh (with a 14 ms refresh interval), while the word remained on the screen on a black background until the participant's vocal response. They were instructed to ignore the word, and to say aloud the colour as quickly as possible into a microphone, without making errors. The experimenter recorded the content of the response (red, green, blue or other) by pressing one of four keys. In the dot probe task, there were 12 practice and 4 bu€er trials, followed by 72 test trials in a new random order for each person. Each trial started with a ®xation cross for 500 ms in the centre of the screen. This was replaced by a word pair in white uppercase letters for 500 ms (approx. 5 mm high), with one word above and one below the central point (words were approximately 3.8 cm apart). Immediately after their o€set, a small dot probe appeared in the location of one of the words, and remained displayed until the participant's response. They 1068 K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 were instructed to press either an upper or lower response key as quickly as possible, while avoiding errors, to indicate the position of the probe. Threat words and probes were presented in the upper or lower position of the display with equal probability. The inter-trial interval varied randomly between 750, 1000 and 1250 ms. After the attentional tasks, questionnaires assessing mood and personality were given including state and trait versions of the State Trait Anxiety Inventory (STAI; Spielberger, Gorsuch, Lushene, Vagg & Jacobs, 1983), Social Desirability Scale (SDS; Crowne & Marlowe, 1964), Beck Depression Inventory (BDI; Beck, Ward, Mendelson, Mock, & Erbaugh, 1961) and ratings of physical and social worries. For the latter, participants rated their negative thoughts and worries on two 9-point scales (ranging from 0=``not at all'' to 8=``extremely'') and were asked to indicate how much they had been recently troubled by (i) anxious thoughts about physical health concerns (e.g., illness, death or accidents), (ii) anxious thoughts about social concerns (e.g., getting on with other people, embarrassment). These worry ratings have been previously used to assess worry±congruent attentional biases in clinical anxiety on both the Stroop and dot probe tasks (e.g., Mogg, Mathews, & Weinman, 1989; Mogg, Mathews, & Eysenck, 1992). 3. Results Alpha level was set at 0.05 level, two-tail, throughout. 3.1. Group characteristics Participants were allocated to three groups according to their scores on STAI trait anxiety and the full version of the SDS (these had been administered under standardised conditions, in contrast to the brief postal, screening measures). The groups were determined according to the following criteria: (1) Low anxious (LA), with trait anxiety scores of less than 37, and SDS less than 15; (2) Repressors (REP), with trait anxiety scores less than 37, and SDS of 15 or more; and (3) High anxiety (HA), with trait anxiety scores of 44 or more, and SDS less than 15. These cuto€s1 produced three approximately equal-sized groups (ns were 16, 14 and 16, respectively), which were similar in age and gender ratio, and also SDS and trait anxiety where appropriate. That is, the HA and LA groups di€ered signi®cantly in trait anxiety, but had comparable SDS scores; while the LA and REP groups di€ered signi®cantly in SDS, but had 1 Division of the sample using normative means as cuto€s (i.e., 40 on trait anxiety, 16 on SDS) produced three unequal-sized groups: 22 LA, 13 REP and 23 HA, with only 2 volunteers not falling into these groups due to above average scores on both measures. Results using these cuto€s were similar to those obtained from the main analyses using more equal-sized groups. That is, for the Stroop task, there was a signi®cant e€ect of group on the interference e€ect of threat (F(2,52)=9.97, p < 0.05); with REP showing less interference of threat than neutral words (ÿ14.0 ms), HA showing greater interference of threat than neutral (7.5 ms), and LA showing no bias (0.5 ms). For the dot probe task, there was a signi®cant e€ect of threat type, but not of group, on bias scores (F(1,54)=4.28, p < 0.05); with greater avoidance of social threat (ÿ6.1 ms) than physical threat (ÿ0.5 ms) words. Mean social threat bias scores were ÿ6.5, ÿ12.8 and ÿ2.3 ms for LA, REP and HA, respectively. Thus, these analyses produced similar ®ndings, but with the disadvantage of unequal sample sizes. 1069 K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 similar trait anxiety scores. Group comparisons using analysis of variance (ANOVA) and Student±Newman±Keuls contrasts (see Table 1a) showed that the HA group had signi®cantly higher trait and state anxiety, Beck depression and social worry scores compared with each of the LA and REP groups, who did not di€er on these measures. The REP group di€ered signi®cantly from the LA group only in SDS scores. The three groups did not di€er in physical worry scores. Fourteen participants did not ful®l the selection criteria for the three groups largely due to having mid-range trait anxiety scores, although their data were included in correlational analyses conducted for the whole sample. 3.2. Preparation of RT data Data from trials with errors were discarded. Box and whisker plots then were used to inspect the RT data from each task and to determine outliers. Consequently, RTs of 200 ms or less were excluded from both tasks, and also those of 900 ms or more in the Stroop task, and 750 ms or more in the dot probe task. Box and whisker plots also indicated that participants' data were outliers if more than 20% were outside the RT cuto€s, or more than 10% were errors, and so these were excluded from the analyses (one HA and one LA were lost from the Stroop task, one REP from the dot probe task). Following these exclusions, the mean percentage of data lost was 3% as errors, and 3% as outliers for the Stroop task; and 2% as errors and 2% as outliers for the dot probe task. The three groups did not di€er signi®cantly in error or outlier rates in either task. The groups also did not di€er signi®cantly in overall mean RT on either task (see Table 1 for mean RTs in each condition). 3.3. Modi®ed Stroop task Bias scores were calculated for each type of threat word by subtracting the mean RT for control words (i.e., categorised neutral) from the mean RT for threat words. This produced two bias scores for each participant: one for social threat, and the other for physical threat words. Positive values of the bias scores indicate a greater interference e€ect of threat words, relative to neutral control words, on colour-naming performance. The mean bias scores of each Table 1a Group characteristics Gender ratio Age Trait anxiety Social desirability State anxiety Beck depression Social worry Physical worry ab LA REP HA F(2,43) p 8/8 20.0 31.6a 9.4a 31.1a 5.6a 2.5a 2.6 7/7 19.9 31.0a 19.8b 31.2a 3.3a 2.6a 2.2 8/8 19.4 52.4b 8.8a 44.8b 13.7b 5.1b 2.3 1.12 67.95 40.95 20.75 20.76 13.76 0.21 ns < 0.001 < 0.001 < 0.001 < 0.001 < 0.001 ns Values with di€erent superscripts di€er at 5% level 2-tail. 1070 K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 Table 1b Mean RTs for each word type condition in Stroop task Social threat Physical threat Categorised neutral Uncategorised neutral LA REP HA 457.1 460.4 457.6 457.6 494.3 492.1 504.9 502.7 483.7 476.1 473.6 474.9 group are shown in the top part of Fig. 1. These scores were entered into a 3  2 mixed design ANOVA with one between-subjects variable of group (3: LA, REP, HA) and one withinsubjects variable of threat type (2: social, physical). This showed a signi®cant main e€ect of group on threat bias scores (F(2,41)=6.05, p < 0.01). The main e€ect of threat type (social vs physical), and the group  threat type interaction, were not signi®cant (Fs < 1). Regarding the main e€ect of group on bias scores, Student±Newman±Keuls contrasts showed that the REP group showed signi®cantly less interference in colour-naming threat words (relative to neutral words) compared with both the LA group (mean bias scores were ÿ12 ms vs 1 ms, p < 0.05) and the HA group (ÿ12 vs 6 ms, p < 0.05). The HA and LA groups did not signi®cantly di€er from each other in threat bias scores. Contrasts of the bias scores against zero (where 0=no bias) indicated signi®cantly less interference of threat than neutral words in the REP group (t(13)=2.39, p < 0.05); and signi®cantly greater interference of threat than neutral words in the HA group (t(14)=2.28, p < 0.05). Speci®c hypothesis-driven analyses were also conducted to test the prediction from Fox (1993) of greater avoidance of social threat words in repressors. The three groups signi®cantly di€ered in social threat bias scores (F(2,41)=5.92, p < 0.01), with the REP group showing signi®cantly less interference of social threat, relative to neutral words, than the HA group (Student±Newman±Keuls, p < 0.05). Contrasts of the social threat bias scores against zero indicated signi®cantly less interference due to social threat than neutral words in the REP group (t(13)=2.37, p < 0.05), and signi®cantly more interference due to social threat than neutral words in the HA group (t(14)=2.36, p < 0.05). Pearson correlations were calculated using the data from the whole sample, including those Table 1c Mean RTs for each condition in dot probe task Threat type Threat position Probe position LA REP HA Social Upper Upper Lower Lower Upper Upper Lower Lower Upper Lower Upper Lower Upper Lower Upper Lower 352.7 355.7 355.1 364.5 356.4 357.8 346.1 350.6 376.3 369.9 356.9 378.0 383.9 364.6 377.9 358.2 381.4 372.0 370.7 373.8 380.3 375.0 377.2 369.5 Physical K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 1071 who did not ful®l the selection criteria for the three groups. The threat bias score correlated signi®cantly and negatively with SDS (r=ÿ0.43, p < 0.01), and positively with trait anxiety (r = 0.27, p < 0.05), state anxiety (r = 0.31, p < 0.05), Beck depression (r = 0.28, p < 0.05), and social worry scores (r = 0.30, p < 0.05). There was no relationship between the threat bias score and physical worry ratings (r=ÿ0.01, ns). The lack of a relationship between physical worry and trait anxiety may be partly due to ¯oor e€ects, as we used a healthy young sample of students who seemed to have a relatively low level of physical health concerns. Fig. 1. Mean attentional bias scores (in ms) for social and physical threat words in low anxiety (LA), repressor (REP) and high anxiety (HA) groups on the modi®ed Stroop and dot probe task. 1072 K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 When the e€ect of SDS scores was controlled using partial correlations, there was a nearsigni®cant trend for the threat bias score to be associated with social worry ratings (r = 0.25, p < 0.07, two-tail). No other correlations approximated signi®cance after controlling SDS (e.g., threat bias and trait anxiety: r = 0.14, ns). However, when the e€ects of trait and state anxiety, Beck depression and social worry scores were partialled out together, the threat bias score remained signi®cantly and negatively correlated with SDS (r=ÿ0.38, p < 0.01). With regard to inter-correlations between the questionnaire measures, SDS scores correlated negatively with trait anxiety (r=ÿ0.34, p < 0.01) and Beck depression (r=ÿ0.46, p < 0.05), but not signi®cantly with the other emotion measures. The emotion measures were all positively and signi®cantly intercorrelated (rs > 0.51, ps < 0.01), with the exception of physical worry scores (e.g., physical worry and trait anxiety: r = 0.02, ns). A supplementary analysis was carried out to examine the e€ect of task order on bias scores (i.e., whether the Stroop task preceded or followed the dot probe task). An ANOVA of bias scores with group (3), threat type (2), and task order (2) as independent variables showed no signi®cant order e€ects (e.g., group  task order, F(2,38)=1.10, ns), so there was no evidence that task order was a confounding factor. To check a subsidiary issue regarding the e€ect of word categorisation on colour-naming performance, an ANOVA was conducted of the RT data for the two neutral word types, with group (3) and categorisation (2: categorised neutral versus uncategorised neutral) as independent variables (see Table 1b for means). This showed no signi®cant results (e.g., main e€ect of word categorisation: F < 1). 3.4. Dot probe task Bias scores were calculated for each type of threat word by subtracting the mean RT when the probe was in the same position as a threat word from the mean RT when the probe was in the opposite location (MacLeod & Mathews, 1988; Mogg et al., 1995). The bias score summarises the interaction e€ect of Threat position  Probe position in the RT data. Positive values of the bias score indicate relatively faster RTs for probes replacing threat rather than neutral words, i.e., an attentional bias towards threat (vigilance); whereas negative values re¯ect a bias away from threat. Two bias scores were calculated for each participant: one for trials with social threat Ð neutral word pairs, and the other for physical threat Ð neutral word pairs. Mean bias scores of each group are shown in the lower part of Fig. 1. These scores were entered into a 3  2 mixed design ANOVA with one between-subjects variable of group (3: LA, REP, HA) and one within-subjects variable of threat type (2: social, physical). The main e€ect of group on bias scores (F < 1) and the interaction e€ect of group  threat type (F(2,42)=1.04, ns) were not signi®cant. However, there was a signi®cant main e€ect of threat type (F(1,42)=5.00, p < 0.05), suggesting generally greater avoidance (i.e., more negative bias scores) for social threat than physical threat words; see Fig. 1. Contrast of the overall mean social threat bias score against zero indicated a signi®cant attentional bias against social threat words (t(44)=3.40, p < 0.01), indicating that on trials with social threat words, participants were signi®cantly faster to detect probes replacing neutral than threat words. There was no evidence of an overall bias either for or against physical threat words (contrast of bias score vs zero: t < 1, ns). K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 1073 Speci®c hypothesis-driven analyses were conducted to examine whether Fox's (1993) ®nding from the dot probe task of avoidance of social threat words in repressors had been replicated. There was a non-signi®cant trend for the three groups to di€er in bias scores for social threat words (F(2,44)=2.08, p = 0.14). Mean social threat bias scores for the LA, REP and HA groups were ÿ1.7 ms, ÿ13.7 ms and ÿ6.3 ms, respectively. Contrasts of the bias scores versus zero indicated signi®cant avoidance of social threat words, relative to neutral words, in the REP group (t(12)=4.80, p < 0.001), whereas the mean social threat bias scores of the LA and HA did not di€er signi®cantly from zero. Correlations between the bias scores from the dot probe task and the questionnaire measures showed a near-signi®cant trend for high SDS scores to be associated with greater avoidance of social threat words (r=ÿ0.23, p = 0.08, two-tail). There were no signi®cant results. A supplementary analysis was carried out to examine the e€ect of task order on bias scores from the dot probe task. An ANOVA with group (3), threat type (2), and task order (2: Stroop before dot probe task, and vice versa) as independent variables showed no signi®cant order e€ects on bias scores (e.g., group  task order, threat type  task order: Fs < 1). 3.5. Relationship between bias scores from modi®ed Stroop and dot probe tasks Pearson correlations showed no signi®cant relationships between the bias scores from the two tasks across the whole sample. The mean threat bias scores from the Stroop task correlated ÿ0.03 with those from the dot probe task. The social threat bias scores from the two tasks correlated 0.01 with each other; and the physical threat bias scores correlated 0.13. 4. Discussion One of the aims of this study was to investigate whether high levels of defensiveness are associated with attentional avoidance of threat (cf. Fox, 1993; Bonanno & Singer, 1990), or with greater vigilance for threat (cf. Dawkins & Furnham, 1989). The results from both cognitive tasks were compatible in indicating that individuals with low levels of trait anxiety and high levels of defensiveness (so-called ``repressors'') had a bias to allocate processing resources away from social threat stimuli. Speci®cally, they were faster in colour-naming social threat than neutral words, and they were faster in detecting probes that replaced neutral rather than social threat words. These particular results were directly predicted from Fox's (1993) ®ndings, and are consistent with the view that high levels of defensiveness combined with low trait anxiety are associated with an avoidant attentional style (although it should be noted that the group di€erence in bias scores on the dot probe task was not signi®cant; a point which we shall return to later). Nevertheless, the present results are contrary to those of Dawkins and Furnham (1989), where repressors showed greater interference in colour-naming threat than neutral words. This discrepancy may be due to methodological di€erences between versions of the Stroop task. In the computerised version used here, words were presented one at a time, and participants' vocal responses terminated the threat display. Thus, an avoidant attentional style (as suggested by faster RTs on trials with threat than neutral words) was e€ective in reducing exposure to 1074 K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 the threat stimuli in the present study. However, on the card version used by Dawkins and Furnham (1989), participants were presented with blocks of 100 words, and an avoidant attentional style would be ine€ective in removing the threat stimuli. Thus, one possibility is that defensive individuals may show enhanced disruption or distraction e€ects of taskirrelevant threat when their attempts to avoid processing the threat are thwarted by the task demands. Another possibility is that Dawkins and Furnham's results are not representative of repressors, given a recent failure to replicate them with a card version of the Stroop task (Myers & McKenna, 1996). In the present study, the two cognitive tasks did not produce identical ®ndings. Signi®cant group di€erences in attentional bias were indicated by the Stroop task, but not by the dot probe task. In the former task, the attentional bias was re¯ected by the di€erence in colournaming times for threat versus neutral words and the results suggested vigilance for threat in high trait anxiety, no bias in low trait anxiety, and avoidance in repressors (as predicted by Eysenck's (1997) theory). Partial correlations indicated that the attentional bias for threat was predicted by SDS scores, independently of anxiety or worry. There was also a near-signi®cant trend for the bias to be predicted by social worry ratings, independently of the e€ect of SDS scores. These results suggested that both worry and defensiveness contribute to the bias for threat words. On the other hand, the dot probe task appeared to be less sensitive to e€ects of anxiety and SDS in this non-clinical sample, given the lack of signi®cant group di€erences. Instead, it indicated greater overall avoidance of social threat than physical threat words, with a nearsigni®cant trend in the correlational results for the former to be associated with increased defensiveness. MacLeod et al. (1986) noted a similar general tendency for their non-clinical sample to be avoidant of threat on the dot probe task, although the speci®city of this e€ect to social versus physical threat words was not noted in their study. The ®ndings from non-clinical studies using the dot probe task with word stimuli have been variable (e.g., review by Mogg & Bradley, 1998), although evidence of trait anxiety-related biases has been obtained, for example, under stressful conditions (e.g., MacLeod & Mathews, 1988), or with brief, masked exposure conditions (e.g., Bradley, Mogg & Lee, 1997). In general, the dot probe task appears to provide a relatively fragile index of anxiety-related attentional biases in non-clinical studies, particularly when using word stimuli that have relatively mild threat value. However, it is reassuring that more persuasive evidence of such biases has been found in non-clinical samples with variants of the dot probe task using more salient stimuli, such as angry faces (e.g., Bradley, Mogg, Falla & Hamilton, 1998; Mogg & Bradley, 1999). One possibility is that, for mild stimuli such as single threat words, high trait anxious individuals may ®nd it easier to use strategies to counteract their vigilant tendencies, which may explain the diculty in reliably demonstrating such vigilance in non-clinical studies on the dot probe task. In the present study, there was no evidence whatsoever of positive relationships between the cognitive bias measures from the dot probe and Stroop tasks. This suggests that performance on each task may be tapping quite di€erent underlying mechanisms. For example, the dot probe task re¯ects biases in visual orienting that depend on the presence of threat cues in di€erent spatial locations in the visual ®eld. Thus, such biases may involve a variety of operations in visual orienting, such as shifting, engagement and disengagement of attention (e.g., Posner, 1995). By contrast, the Stroop task involves competition in processing di€erent K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 1075 attributes of a single stimulus presented within the focus of attention (i.e., colour versus emotional content of word). Using the analogy of a ``spotlight of attention'', the dot probe task may re¯ect the e€ect of threat cues on the scanning of the spotlight across di€erent regions of the visual ®eld, whereas the Stroop task involves a stationary spotlight where the information at the focus of its beam must be disentangled. The present results highlight the importance of clarifying the precise mechanisms underlying biases on various cognitive tasks. Moreover, the speci®c mechanisms involved in the avoidant attentional style of defensive individuals may not necessarily be the same as those underlying the vigilant attentional style of generally anxious patients. For example, such avoidant and vigilant biases may vary in the extent to which they involve shifting versus disengagement of attention, or the extent to which they are automatic versus amenable to strategic control. Despite the absence of signi®cant relationships between the di€erent measures of attentional bias, the results from the two tasks were compatible in suggesting not only that low trait anxious individuals with high levels of defensiveness showed attentional avoidance of social threat (as noted above), but also that low trait anxiety without high defensiveness was generally associated with an unbiased attentional style (see Fig. 1). Both of these ®ndings were predicted by Eysenck's (1997) cognitive theory of anxiety, which proposed that attentional biases for threat-related information are in¯uenced by more than one personality variable, namely, anxiety-proneness and defensiveness. The present results also highlight the importance of including a measure of defensiveness in research into anxiety-related cognitive biases. If this had not been included in the present study, analyses of the Stroop results by separating into low and high trait anxiety groups would have suggested a marked e€ect of trait anxiety on the processing of threat. This would seem to be a misleading interpretation of the data, which in fact suggested instead that defensiveness (as re¯ected by SDS scores) was the more important variable predicting the threat interference e€ect in this non-clinical sample. Indeed, no signi®cant di€erences in attentional bias were found between the high and low trait anxious groups that had similar SDS scores. The main evidence for an anxiety-related bias was a notable (near-signi®cant) trend for social worries to predict threat interference, independently of defensiveness. Inclusion of a measure of defensiveness in future research may therefore prove helpful in accounting for discrepant ®ndings across non-clinical studies. On a more general note, it would seem preferable to use the term ``defensive, low anxious'' rather than ``repressor'' to refer to these individuals, given that this more objectively re¯ects the criteria used in selection. Indeed, one reason why researchers may be reluctant to include measures of defensiveness in their work is unease with the term ``repression'' and its Freudian origins, and so more neutral (i.e., assumption-free) terminology would seem advantageous. The study of the e€ects of personality variables, such as defensiveness and trait anxiety, on information processing biases is not only of theoretical interest, but also has potentially important clinical implications. For example, it is unclear whether attentional avoidance strategies (as shown by defensive, low anxious individuals or ``repressors'') are protective against clinical anxiety, and whether it is helpful to provide attentional retraining for anxious patients to encourage them to adopt such an avoidant attentional style. Such clinical interventions are indeed suggested by Williams et al.'s (1997) cognitive model of emotional disorders, which proposes that attentional vigilance for threat is a key variable underlying vulnerability to clinical anxiety (but see Mogg & Bradley, 1998, for further debate). 1076 K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 In summary, the present results suggest that high levels of defensiveness combined with low trait anxiety were associated with an avoidant attentional style in this non-clinical sample. Our ®ndings also indicate that research into cognitive models of non-clinical and clinical anxiety should not only routinely include a measure of defensiveness, but it should also seek convergent evidence from more than one index of attentional bias, given that di€erent tasks are likely to re¯ect di€erent underlying attentional mechanisms. Acknowledgements This work was supported in part by the Wellcome Trust. Karin Mogg holds a Wellcome Trust Senior Research Fellowship in Basic Biomedical Science. References Beck, A. T., Ward, C. H., Mendelson, M., Mock, J., & Erbaugh, J. (1961). An inventory for measuring depression. Archives of General Psychiatry, 4, 561±571. Bendig, A. W. (1956). The development of a short form of the Manifest Anxiety Scale. Journal of Consulting Psychology, 20, 384. Bonanno, G. A., & Singer, J. L. (1990). Repressive personality style: Theoretical and methodological implications for health and pathology. In J. L. Singer, Repression and dissociation. Chicago: University of Chicago Press. Bonanno, G. A., Davis, P. J., Singer, J. L., & Schwartz, G. E. (1991). The repressor personality and avoidant information processing: A dichotic listening study. Journal of Research in Personality, 25, 386±401. Bradley, B. P., Mogg, K., & Lee, S. (1997). Attentional biases for negative information in induced and naturally occurring dysphoria. Behaviour Research & Therapy, 35, 911±927. Bradley, B. P., Mogg, K., Falla, S. J., & Hamilton, L. R. (1998). Attentional bias for threatening facial expressions in anxiety: e€ect of stimulus duration. Cognition & Emotion, 12, 737±753. Bradley, B. P., Mogg, K., Millar, N., Bonham-Carter, C., Fergusson, E., Jenkins, J., & Parr, M. (1997). Attentional biases for emotional faces. Cognition & Emotion, 11, 25±42. Carroll, J. B., Davies, P., & Richman, B. (1971). Word frequency book. New York: American Heritage. Crowne, D. P., & Marlowe, D. (1964). The approval motive: studies in evaluative dependence. New York: Wiley. Dawkins, K., & Furnham, A. (1989). The colour naming of emotional words. British Journal of Psychology, 80, 383±389. Derakshan, N., & Eysenck, M. W. (1997). Interpretive biases for one's own behaviour and physiology in high trait anxious individuals and repressors. Journal of Personality and Social Psychology, 73, 1±10. de Ruiter, C., & Brosschot, J. F. (1994). The emotional Stroop interference in anxiety: attentional bias or cognitive avoidance. Behaviour Research & Therapy, 32, 315±319. Eysenck, M. W. (1992). Anxiety: the cognitive perspective. Hove: Erlbaum. Eysenck, M. W. (1997). Anxiety and cognition: a uni®ed theory. Hove: Psychology Press. Fischer, D. G., & Fick, C. (1993). Measuring social desirability: short forms of the Marlowe±Crowne Social Desirability Scale. Educational and Psychological Measurement, 53, 417±424. Fox, E. (1993). Allocation of visual attention and anxiety. Cognition & Emotion, 7, 207±215. MacLeod, C. M. (1991). Half a century of research on the Stroop e€ect: An integrative review. Psychological Bulletin, 109, 163±203. MacLeod, C., & Mathews, A. (1988). Anxiety and the allocation of attention to threat. Quarterly Journal of Experimental Psychology, 40, 653±670. MacLeod, C., Mathews, A., & Tata, P. (1986). Attentional bias in emotional disorders. Journal of Abnormal Psychology, 95, 15±20. K. Mogg et al. / Personality and Individual Di€erences 28 (2000) 1063±1077 1077 Mathews, A., & MacLeod, C. (1994). Cognitive approaches to emotion and emotional disorders. Annual Review of Psychology, 45, 25±50. Mogg, K., & Bradley, B. P. (1998). A cognitive-motivational analysis of anxiety. Behaviour Research & Therapy, 36, 809±848. Mogg, K., & Bradley, B. P. (1999). Some methodological issues in assessing attentional biases for threatening faces in anxiety: a replication study using a modi®ed version of the probe detection task. Behaviour Research & Therapy, 37, 595±604. Mogg, K., Bradley, B. P., Williams, R., & Mathews, A. (1993). Subliminal processing of emotional information in anxiety and depression. Journal of Abnormal Psychology, 102, 304±311. Mogg, K., Bradley, B. P., & Williams, R. (1995). Attentional bias in anxiety and depression: the role of awareness. British Journal of Clinical Psychology, 34, 17±36. Mogg, K., Mathews, A., & Eysenck, M. (1992). Attentional bias in clinical anxiety states. Cognition & Emotion, 6, 149±159. Mogg, K., Mathews, A., & Weinman, J. (1989). Selective processing of threat cues in clinical anxiety states: a replication. Behaviour Research & Therapy, 27, 317±323. Myers, L. B., & McKenna, F. P. (1996). The colour naming of socially threatening words. Personality and Individual Di€erences, 20, 801±803. Posner, M. I. (1995). Attention in cognitive neuroscience: an overview. In Gazzaniga, The cognitive neurosciences. Cambridge, MA: MIT Press. Posner, M. I., Snyder, C. R., & Davidson, B. J. (1980). Attention and the detection of signals. Journal of Experimental Psychology: General, 109, 160±174. Schneider, W. (1988). Micro Experimental Laboratory: An integrated system for IBM PC compatibles. Behaviour Research Methods, Instruments & Computers, 20, 206±217. Spielberger, C. D., Gorsuch, R. L., Lushene, R., Vagg, P. R., & Jacobs, G. A. (1983). Manual for the State-Trait Anxiety Inventory. Palo Alto, CA: Consulting Psychologists Press. Strahan, R., & Gerbasi, K. C. (1972). Short, homogeneous versions of the Marlowe±Crowne social desirability scale. Journal of Clinical Psychology, 28, 191±193. Weinberger, D. A., Schwartz, G. E., & Davidson, J. R. (1979). Low-anxious, high-anxious, and repressive coping styles: Psychometric patterns and behavioural and physiological responses to stress. Journal of Abnormal Psychology, 88, 369±380. Williams, J. M. G., Watts, F. N., MacLeod, C., & Mathews, A. (1997). Cognitive psychology and emotional disorders (2nd ed.). Chichester: Wiley.