Behavior Therapy 37 (2006) 237–247www.elsevier.com/locate/bt

Positive Interpretation Training: Effects of Mental Imagery VersusVerbal Training on Positive Mood

Emily A. Holmes, Andrew Mathews, Tim DalgleishMRC Cognition and Brain Sciences Unit, Cambridge, UK

Bundy MackintoshMRC Cognition and Brain Sciences Unit, Cambridge, UK, and Open University, Milton Keynes, UK

Therapists often assume a special association betweenmental imagery and emotion, though empirical evidencehas been lacking. Using an interpretation training paradigm,we previously found that imagery had a greater impact onanxiety than did verbal processing of the same material(Holmes & Mathews, 2005). Although the finding of adifferential impact of imagery versus verbal processing ofnegative material was replicated, findings did not extend tobenign material. Results therefore left open the question ofwhether there may be a special association between imageryand positive emotion. The current experiment examinedpositive interpretation training. Numerous scenarios werepresented with initial ambiguity as to positive outcome ornot, with final information then yielding consistently positiveresolutions. Participants were asked to either imagine thesepositive events or to listen to the same descriptions whilethinking about their verbal meaning. Those participants inthe imagery condition reported greater increases in positiveaffect and rated new descriptions as beingmore positive thandid those in the verbal condition. Results suggest thatpositive training can be enhanced through imagery asopposed to verbal processing. This study also provides thefirst test of a standardized intervention using an “interpretivebias training” paradigm to improve positive mood.

Emily Holmes is now at the University of Oxford, Oxford, UK.Andrew Mathews is now at the University of California at Davis.

Address correspondence to Dr. Emily Holmes, Royal SocietyDorothy Hodgkin Fellow, University of Oxford, Department ofPsychiatry, Warneford Hospital, Oxford, OX3 7JX, UnitedKingdom; e-mail: emily.holmes@psych.ox.ac.uk.0005-7894/06/0237–0247$1.00/0© 2006 Association for Behavioral and Cognitive Therapies. Published byElsevier Ltd. All rights reserved.

IN COGNIT IVE BEHAVIORAL THERAP IES , relevantcognitive processes include both verbal thoughtsand mental images. For example, distressingintrusive mental images have been reported acrossa range of psychological disorders, includingposttraumatic stress disorder, agoraphobia, socialphobia, body dysmorphic disorder, and psychosis(Holmes & Hackmann, 2004). There has been along-held assumption in both clinical and experi-mental psychology that mental imagery has aprivileged relationship to emotion (Holmes &Mathews, 2005). However, until recently therewas little empirical evidence to support the as-sumption that mental images are associated withmore emotion than are verbal thoughts about thesame material.Holmes and Mathews (2005) used an interpre-

tation training paradigm to compare the processingof text stimuli with imagery versus verbal instruc-tions. In this paradigm, the method used wassimilar to that used in other studies designed toinvestigate induced biases in emotional interpreta-tion (Grey & Mathews, 2000; Mathews &Mackintosh, 2000; Mathews & MacLeod, 2002).Many everyday events are ambiguous and can beinterpreted in a negative, benign, or positive way.For example, a friend walking past withoutacknowledgment might be assumed to be ignoringyou (negative interpretation) as opposed to simplybeing preoccupied (benign interpretation). Negative(rather than benign or positive) interpretationbiases are often thought to be an underlyingcognitive mechanism in both anxiety and mooddisorders. Experimental modification of such biasesthus represents a promising new method that mayfind future application in the clinical domain. That

238 holmes et al .

is, similar methods can perhaps be developed forclinical use to modify underlying negative interpre-tation biases in a more positive direction.In some previous interpretation training experi-

ments (e.g., Mathews & Mackintosh, 2000),participants were asked to read about andimagine themselves in numerous situations.Depending on assignment to experimental group,the descriptions of each situation constrainedparticipants to resolve the outcome of thesituation (which initially appeared ambiguous) ineither a negative or benign way. Results of theseprior studies showed that, compared to partici-pants in the benign condition, those in thenegative resolution condition interpreted newambiguous events more negatively and sometimesreported greater increases in state anxiety. Theseresults supported the hypothesis of a causal linkbetween interpretative bias and anxiety but didnot provide a test of whether the instructions toimagine the situations were critical in producingemotional effects.Holmes and Mathews (2005) contrasted imagery

versus verbal processing instructions using a similarinterpretation training method, with the novelmodification that scenarios were presented inauditory form rather than as written text. In aninitial experiment, participants either imaginedunpleasant events (i.e., used imagery) or listenedto descriptions of the same events while thinkingabout their meaning in verbal terms. Those in theimagery condition reported more anxiety and ratednew ambiguous test descriptions as more emotionalthan did those in the verbal condition. This resultsuggested the possibility that type of processing cancarry over and influence how new descriptions areencoded.In a second experiment, four groups listened to

either benign or unpleasant descriptions, again withimagery or verbal processing instructions. Anxietyagain increased more after unpleasant (but notbenign) imagery than after verbal processing;however, emotionality ratings for new ambiguousdescriptions did not differ between groups after a10-min filler task. This result suggested that theassociated finding in Experiment 1 may have beendue to mood. Overall, the findings of these twoexperiments provided support for the hypothesisthat imagery of negative material has greater effectson self-reported anxiety than does verbal proces-sing of the same material. In contrast, there was noevidence of a parallel effect of imagery versus verbalprocessing with more benign material (i.e., signif-icantly greater reductions of negative affect in thebenign imagery condition), although the meanscores were in the predicted direction.

There are a number of possible explanationsfor this null finding with benign material. Theseinclude limitations in the measure of emotionemployed—the “state” score of the State-TraitAnxiety Inventory (STAI; Spielberger, Gorsuch,Lushene, Vagg, & Jacobs, 1983)—which was notdesigned to assess positive emotions. Further-more, the benign descriptions were often notexplicitly positive. The benign training materialwas designed to be identical to the negativetraining descriptions up until the last word (thatresolved the ambiguity). This structure tended tohave the effect of making the benign descriptionsnonnegative, rather than overtly positive. Theupshot was that the benign material usedprobably did not provide an optimal test ofwhether imagery can produce greater positiveemotional effects than verbal processing.The results of Holmes and Mathews (2005)

therefore supported the clinical assumption thatthere may be a special link between imagery andanxiety but left open the question of whether thisconclusion also applies to other emotions, includingpositive affect. In clinical settings it is important toknow whether positive emotions can be enhanced,for example, by modifying cognitive biases such asthose involved in positive emotional interpreta-tions. For this reason, the present experiment wasdesigned to extend the findings reported in Holmesand Mathews (2005) by comparing interpretationtraining using imagery versus verbal processing ofdescriptions that could be resolved in a more clearlypositive manner and by adding a better measure ofpositive affect. The current experiment thus includ-ed the total positive affect score from the Positiveand Negative Affect Schedule (PANAS; Watson,Clark, & Tellegen, 1988) in addition to the STAImeasure of anxiety. One hundred new positivetraining paragraphs were also created for thisexperiment in order to employ truly positive (ratherthan benign) training material.The results of the second experiment by Holmes

and Mathews (2005) failed to replicate one effectfound earlier, that prior practice in imagery versusverbal processing increased emotionality ratings fornew ambiguous test descriptions when they werepresented after an intervening task. One possibleexplanation for this difference across experiments isthat later emotional effects depend on spontaneousdeployment of the practiced processing style(Hertel, 2002) and that this deployment wasdisrupted in the imagery condition by the verbalfiller task. The current experiment therefore used afiller task designed to be less likely to discourageprocessing using mental imagery—listening tomusic and making pleasantness ratings. With this

239mental imagery and pos i t i v e interpretat ion tra in ing

new filler task, we predicted that imagery trainingwould have a more positive effect on emotionalityratings of ambiguous test scenarios than wouldverbal training.In summary, the key hypotheses to be tested were

that (a) after positive interpretation training,participants in the imagery condition would showa greater increase in positive affect (on the PANAS)than would those in the verbal condition, withcomplementary reductions in anxiety (on the STAI);and (b) interpretive bias, assessed using emotional-ity ratings of ambiguous test scenarios (bothpositively and negatively resolvable) given pretrain-ing and 10 min posttraining, would become morepositive in participants given positive imagerytraining than in those given positive verballyfocused training.

Methodoverview

In positive interpretation training, 100 scenarioswere presented that, although initially ambiguousas to their outcome, subsequently yielded consis-tently positive resolutions. Participants were askedto either imagine these positive events or to listen tothe same descriptions while thinking about theirverbal meaning. A mixed design was used, in whichparticipants were randomly allocated to eitherimagery or verbal processing conditions, withboth using the same positive interpretation trainingmaterials. A state anxiety measure and a positiveaffect measure were completed pretraining, imme-diately posttraining, and after a further 10-minfiller task. Emotionality ratings of two sets ofrepeated ambiguous test paragraphs (one set withpotential negative resolutions and another withpotentially positive resolutions) were completedpretraining and after the filler task. Participantswere debriefed using questions designed to assessexperimental demand and the effectiveness of themanipulation.

participants

The 26 participants comprised 17 women and 9men, with a mean age of 38.85 years (SD = 15.64).They were recruited through the department volun-teer panel (a group of 2,000 community volunteers)and were paid a small fee for participation.

materials

Positive training paragraphs. One hundred descrip-tions were created for the experiment, each describ-ing a situation having a positive emotional outcome.These descriptions were read aloud in the samefemale voice (each lasting approximately 10 to 13 s)

and were digitally recorded using Cool Edit 2000software (Phoenix; Syntrillium Software Corpora-tion). They were presented stereophonically viaheadphones, using E-Prime software (Version1.1.4.1,Pittsburgh;PsychologySoftwareTools Inc.).The structure of the paragraphs was designed to

be similar to those used in previous experiments inthat the positive outcome only became clear in thelast few words. Half the paragraphs began with apotentially negative situation being implied butwere resolved in the final word(s) to have a benignor positive outcome. The remaining paragraphsbegan with a benign situation and ended even morepositively. For example, one description of theformer type (possible negative resolved as benign)read as follows: “You are at home alone watchingTV. You must have been dozing because yousuddenly wake up. You have the impression thatyou heard a frightening noise and then realize withrelief that it was your partner returning home”(resolution in italics). Note that the initial part ofthe scenario was designed to be ambiguous in thesense that it could imply a negative outcome (suchas the noise being due to an unwanted intruder). Anexample of the second type (benign resolved aspositive) was “It’s your birthday, and your partnerreaches over to you with a present. You open it andfeel incredibly happy” (resolution in italics). Despitebeing generally positive, the initial part of thescenario was still intended to be somewhat ambig-uous in the sense that it could be resolved byoutcomes less positive than “feel incredibly happy,”such as “feel pleased” or even “feel disappointed.”All the scenarios used had more than one possibleoutcome, and the aim of using the above structurewas to train participants to generate positiveresolutions of situations that could have developedin other and less desirable ways. The items wererandomized throughout five training blocks, eachof 20 paragraphs, within a program written usingE-prime.In order to focus participants on their assigned

task, after each training paragraph they either ratedvividness of imagery (“How vividly could youimagine the situation that was described?”) or theirability to comprehend the description (“Howdifficult was it to understand the meaning of thedescription?”), depending on condition (imageryvs. verbal). Both ratings were made on a 7-pointscale (1 = not at all and 7 = very). In addition,reminders of the task instructions were givenbetween blocks of 20 training paragraphs.

Filler items given after training. During the 10-min interval after training, participants performedan unrelated filler task of listening to music. Thereason for including this interval was that we

240 holmes et al .

wanted participants’ mood to return to baselinelevels prior to the posttraining test paragraphs inorder to minimize the influence of any mooddifferences across groups on interpretation ofthese paragraphs. For the filler task, a series ofclassical music extracts was played, each extractlasting 40 s. Participants were asked to rate eachextract for how pleasant they found it on a scale of1 (extremely unpleasant) to 9 (extremely pleasant).

Ambiguous test descriptions. Twenty descrip-tions were used to test for generalization to newmaterial and were presented both before and afterthe training phase, but without specific instructionsas to encoding. These descriptions were ambiguousin that possible emotional outcomes were impliedbut not explicitly described. The same set of 10negatively resolvable test paragraphs and 10positively resolvable paragraphs were given as inHolmes and Mathews (2005). After each descrip-tion participants were asked to rate “How emo-tional is this description?” using a 9-point scalefrom 1 (extremely unpleasant) to 9 (extremelypleasant). For example, one negatively resolvabledescription read as follows: “You are on a safety-at-work training day. They are showing videos ofworking in a factory. There is a shot of someoneusing a large sawmachine to demonstrate the use ofa safety guard. The shot pans in on the person’sarm, which is highlighted.” An example of apositively resolvable test description read “Youbuy a new outfit for a party. Other people’sreactions show how you look.”

Questionnaire measures. The STAI (Spielbergeret al., 1983) was used to measure trait and stateanxiety. The STAI Trait scale consists of 20 anxiety-related items for which participants rate “how yougenerally feel” on a 4-point scale: almost never,sometimes, often, or always. The STAI State scaleconsists of 20 anxiety-related items for whichparticipants rate how they feel “right now, that isat this moment” on a 4-point scale: not at all,somewhat, moderately so, or very much so. Thesewidely used measures are reported to have satisfac-tory reliability and validity (Spielberger et al.,1983). For example, administration of the STAIState scale to samples of working adults and collegestudents have reported alpha coefficients of above0.90, indicating good internal consistency.State positive affect was measured using all of the

positive affect subscales of the PANAS (Watson etal., 1988) to calculate the total positive affect score.Negative affect items were excluded. The positivesubscales include a total of 21 items, divided intothe basic positive emotion scales (joviality, 8 items;self-assurance, 6 items; attentiveness, 4 items) aswell as the serenity subscale (3 items), as detailed by

Watson and Clark (1994). As these subscales wereto be used as a repeated measure of state affect, theitems were administered with the short-term timeinstructions (“Indicate to what extent you feel thisway now / in the past few minutes”). Watson andClark (1994) report that these subscales aresensitive to changing internal or external circum-stances such as social activity, physical exercise, anddiurnal variation.The tendency to use imagery in everyday life was

measured using the Spontaneous Use of ImageryScale (SUIS; Reisberg, Pearson, & Kosslyn, 2003).This questionnaire has 12 items, for example,WhenI think about visiting a relative, I almost alwayshave a clear mental picture of him or her and BeforeI get dressed to go out, I first visualize what I willlook like if I wear different combinations of clothes.Each item is rated on a 5-point scale, anchored withthe following instructions: “If a description isalways completely appropriate, please write 5; if itis never appropriate, write 1; if it is appropriateabout half of the time, write 3; and use the othernumbers accordingly.” Reisberg et al. (2003) foundthat the mean score (average across all items) for150 participants was 3.1, with a range of 1.2 to 4.7.These authors also found a significant relationshipbetween scores on the SUIS and the Vividness ofVisual Imagery Questionnaire (VVIQ; Marks,1973), with high-vividness imagers reporting higheruse of imagery than low-vividness imagers. Theauthors have classified mean scores below 2.5 aslow imagery use and anything above 3.5 as highimagery use (D. G. Pearson, personal communica-tion, September 23, 2003).

Subjective experience ratings. Questions werealso given to assess the extent to which participantsreported using imagery or verbal processing duringthe training phase and during the test phase. Eachquestion was rated on a 9-point scale, ranging from0 (not at all) to 8 (all the time).

Demand questions (impact predictions). Partici-pants were asked to rate their expectation abouthow participating in each condition might influencetheir emotions (Baddeley & Andrade, 2000). Oneimpact prediction was for the imagery conditionand one for the verbal condition, yielding twoscores (condition impact predictions) per partici-pant. The questions were:Howmuch, if any, wouldyou predict that imaging the sentences you heardrather than listening to them normally, would affectyour feelings? and How much, if any, would youpredict that focusing on the words and meaning ofthe sentences you heard rather than listening tothem normally, would affect your feelings? Bothimpact prediction ratings were made on a 21-pointscale from −10 (very much increase negative

241mental imagery and pos i t i v e interpretat ion tra in ing

feelings) to +10 (very much increase positivefeelings).

procedure

After giving their informed consent to the experi-ment, participants were randomly assigned to eitherthe imagery or verbal condition. They first com-pleted the STAI, SUIS, and PANAS. Then they puton headphones and listened to the first test set of 20ambiguous descriptions, presented in randomorder; immediately after each description, theyrated it for emotionality as described above (seeambiguous test descriptions).The experimenter then read out instructions for

the assigned condition. In the imagery condition,participants were given a brief practice task inwhich they were asked to imagine cutting a lemonin order to clarify what was meant by “using mentalimagery.” They were then given four (nonemotion-al) example descriptions and asked to imagine eachevent as happening to themselves while describingtheir mental image out loud, with a final exampleadministered using the computer. The experimenterexplained that maintaining a focus on their imageswould help in answering the questions thatfollowed.In the verbal condition, participants were

instructed to focus on the meaning of eachdescription as they heard it. They were given apractice sentence about cutting a lemon in order toillustrate the task, and then four nonemotionaldescriptions as practice, with instructions to “con-centrate on the words and meaning as thedescription unfolds.” A final example was admin-istered using the computer. The experimenterexplained that focusing on the words and meaningsof each description would help them to answer thequestions that followed.Participants were then given the 100 training

descriptions in 5 randomized blocks of 20 each,with breaks allowed between blocks. For anyparticipant, depending on their assigned condition,all descriptions were followed either by a vividnessrating or a rating of confidence in their ability tounderstand the verbal content. The STAI andPANAS were repeated at the end of the trainingphase.An interval of 10 min was allowed after the end

of the imagery or verbal training phase and filled bya neutral task (as described above). The emotionmeasures (STAI and PANAS) were then readminis-tered at the end of the interval to check whethermood had returned to baseline. All participantsthen completed a second administration of the 20ambiguous test descriptions, making the sameemotionality ratings as before training. They then

completed the subjective experience ratings toassess reported use of imagery and verbal proces-sing during the training and test phase. Finally, theycompleted two ratings to assess perceived demand(impact predictions). Participants were thendebriefed and thanked for their participation.

ResultsOur key hypotheses were that practice usingimagery, relative to verbal processing, wouldproduce greater increases in positive affect andpositive interpretations and reduce anxiety scores.Because these hypotheses were specific and direc-tional, analyses that directly tested them employedone-tailed tests. All other tests, when no directionalhypothesis had been stated, used two-tailed tests.

comparison of participants inimagery and verbal conditions

To check the comparability of participants in theimagery and verbal conditions, the groups werecompared on the background measures (usingindependent samples t-tests unless otherwisereported). These comparisons indicated that therewas no significant difference between the twoconditions in terms of gender, Fisher’s exact test,p = .69. There were 4 men and 9 women in theimagery condition, and 5 men and 8 women in theverbal condition. The two conditions did not differsignificantly according to age, trait anxiety, initialstate anxiety scores, initial positive affect scores, theSUIS imagery questionnaire, emotionality ratings ofnegatively resolvable scenarios pretraining (for eachmeasure, t[24] < 1.0), except for emotionalityratings of positively resolvable scenarios pretrain-ing, t(24) = 1.63, p = .12. The mean age was 36.92(SD = 16.55) in the imagery condition and 40.77(SD = 15.09) in the verbal condition. Mean traitanxiety scores (STAI) were 33.46 (SD = 7.28) in theimagery condition and 31.77 (SD = 5.26) in theverbal condition. Table 1 shows means for pre-training state anxiety and positive affect scores, aswell as for pretraining emotionality ratings ofnegatively and positively resolvable scenarios.

state anxiety pre- and immediatelyposttraining

We predicted that participants in the imagerycondition would show greater reductions in stateanxiety after training than would those in the verbalcondition. This hypothesis was tested using a mixedmodel ANOVAwith a grouping factor of condition(imagery versus verbal) and a within-subjects factorof time (pre- versus posttraining). For mean changescores per condition, see Figure 1. There was nomain effect of time, F(1, 24) = 0.12, p = .73, and no

FIGURE 1 Mean changes in state anxiety and positive affectafter positive imagery and positive verbal training. Error bars showone standard error of the mean.

Table 1Means and standard deviations for the mood measures (STAIand PANAS), emotionality ratings for ambiguous testdescriptions, and impact predictions

Measure Imagerycondition

Verbalcondition

n = 13 n = 13

M SD M SD

Mood measuresState STAI, Time 1 29.77 10.07 30.46 15.64State STAI, Time 2 25.38 4.50 33.85 13.96State STAI, Time 3 26.77 4.57 31.08 12.23PANAS, Time 1 67.62 10.67 70.92 15.42PANAS, Time 2 74.77 8.42 64.84 13.70PANAS, Time 3 70.23 12.48 68.84 13.36Emotionality ratings(descriptions)Negatively resolvable, Time 1 3.46 1.06 3.25 0.80Negatively resolvable, Time 3 3.38 0.55 3.27 0.62Positively resolvable, Time 1 6.43 0.93 6.91 0.48Positively resolvable, Time 3 6.56 0.83 6.24 0.71Impact predictionsImagery task 4.31 2.81 2.69 4.49Verbal task 1.15 1.82 0.62 4.05

Note. Time 1 = pre training, time 2 = immediately post training, andtime 3 = after 10-min filler task post training. STAI = State-TraitAnxiety Inventory, PANAS = total positive affect score from thePANAS. Emotionality ratings for ambiguous test descriptions,1 = extremely unpleasant to 9 = extremely pleasant. For taskimpact predictions (compared to having no task), −10 = very muchincrease negative feelings, 0 = do nothing, +10 = very muchincrease positive feelings.

242 holmes et al .

main effect of condition, F(1, 24) = 1.07, p = .31.However, there was a significant interaction of timewith condition, F(1, 24) = 7.36, p = .012, η2 = .24,consistent with our hypothesis of greater reductionin anxiety following imagery training than follow-ing verbal training.Our directional hypothesis was tested by

comparing the reduction in anxiety scores seenin imagery with those in verbal processinggroups using an independent samples t-test.This analysis confirmed the expected effect,with greater reductions seen in the imagerygroup, mean change = −4.38 (SD = 8.37) vs.+3.38 (SD = 6.04), t(24) = 2.71, p = .006, (one-tailed), d = 1.06. However, because this contrastcould reflect an increase of anxiety in the verbalgroup as well as a decrease in the imagery groupalone, we also tested these changes individually.On two-tailed paired-sample t-tests, there werenonsignificant trends for both a reduction inanxiety after imagery training, t(12) = 1.89,p = .083, d = 0.54, and an increase in anxiety inthe verbal condition, t(12) = 2.02, p = .067,d = 0.23.

positive affect pre- and immediatelyposttraining

The hypothesis that participants in the imagerycondition would show greater increases inpositive affect than those in the verbal conditionwas tested using a similar ANOVA to thatdescribed above (for mean change scores percondition, see Figure 1). There was no maineffect of time, F(1, 24) = 0.08, p = .79, and nomain effect of condition, F(1, 24) = 0.56, p = .46.However, as with anxiety, there was a significantinteraction of time with condition, F(1,24) = 11.43, p = .002, η2 = .32, that wasconsistent with our hypothesis of a greaterincrease in positive affect following imagerytraining as compared to verbal training.We again tested our directional hypothesis by

comparing changes over time for the two groups,confirming that the imagery group reported a largerincrease in positive affect, mean change = +7.15(SD = 10.30) vs. −6.08 (SD = 9.63), t (24) = 3.38,p = .001, one-tailed, d = .11 Again, because thisresult may have been due in part to an unexpecteddecrease in positive affect in the verbal group,changes in each group were analyzed separately.Both changes were significant: the imagery groupshowed a significant increase in positive affect,t(12) = 2.50, p = .028, d = 0.74, and positiveaffect significantly decreased in the verbal condi-tion, t(12) = 2.27, p = .042, d = 0.42, (both two-tailed). These results indicate that, as predicted,participants in the imagery condition showedgreater increases in positive affect after trainingthan those in the verbal condition (see Figure 1).In addition, the verbal group showed an unpre-dicted significant decrease in positive affect.

243mental imagery and pos i t i v e interpretat ion tra in ing

state anxiety and positive affectafter the filler task

Participants were given a 10-min filler task after thetraining phase. The STAI-state and PANAS (posi-tive affect subscales) were then administered for thethird time. The aims were (a) to test whetherparticipants' state anxiety reports had returned tobaseline level and (b) to check that there were noremaining differences between groups in stateanxiety prior to administration of emotionalityratings for the ambiguous paragraphs.

Return to baseline levels. Any differences be-tween mood scores after the 10-min break withbaseline scores were tested using mixed modelANOVAs. These had one within-subjects factor oftime (pretraining versus after the filler task) and abetween-subjects factor of training condition (im-agery versus verbal). Mean mood scores in eachcondition are shown in Table 1.For state anxiety, there were no main effects of

time, F(1, 24) = 0.76, p = .39, nor of trainingcondition, F(1, 24) = 0.35, p = .56. The interactionbetween time and training condition was also notsignificant, F(1, 24) = 1.74, p = .20, η2 = .07.Similarly, for positive affect, there were no maineffects of time, F(1, 24) = 0.03, p = .86, nor trainingcondition, F(1, 24) = 0.04, p = .85. Again, theinteraction between time and training conditionwas not significant, F(1, 24) = 2.38, p = .14,η2 = .09.

Group differences after the filler task. Differ-ences between groups in mood prior to theemotionality ratings test were examined using aone-way ANOVA, with the between-subjects factorof training condition (imagery versus verbal). Therewere no significant differences between groups ineither state anxiety, F(1, 24) = 1.42, p = .25,η2 = .06, or positive affect, F(1, 24) = 0.08, p = .79(for mean scores see Table 1). This result indicatesthat there were no significant differences in moodbetween the two groups after the 10-min interpo-lated task.

change in emotionality ratings forrepeated ambiguous test descriptions

Participants rated the emotionality of the two setsof 10 ambiguous paragraphs both before trainingand then 10 min after training. Set 1 consisted ofnegatively resolvable ambiguous paragraphs (neg-atively biased) and Set 2 of positively resolvableambiguous paragraphs (positively biased). Since thedata from the two sets sampled different valencedomains, the negatively and positively resolvableambiguous paragraphs were analyzed separately.Our second hypothesis was that, compared toparticipants given verbal training, those given

imagery training would rate scenarios overall asmore positive after training.

Negatively resolvable ambiguous paragraphs.Changes in emotionality ratings from pretrainingto posttraining were examined using a mixed modelANOVA in a similar analysis to that for stateanxiety scores. There was no significant main effectof time, F(1, 24) = 0.06, p = 0.81, or condition, F(1,24) = 0.37, p = .56. There was no interactionbetween time and training condition, F(1,24) = 0.16, p = .69. Overall, the results thereforefailed to support a differential effect of imageryversus verbal encoding conditions on emotionalratings for the negatively biased descriptionsencountered after the filler task.

Positively resolvable ambiguous paragraphs.The same analysis as before revealed a nonsignif-icant trend for the main effect of time, F(1,24) = 3.71, p = .066, η2 = .13, but no main effectof condition, F(1, 24) = 0.08, p = .77. However,there was a significant interaction between time andtraining condition, F(1, 24) = 8.09, p = .009,η2 = .25, that was consistent with our hypothesisthat emotional ratings should change more in apositive direction following imagery than followingverbal training.Testing our directional hypothesis as before we

compared the changes on ratings of positivelyresolvable items before and after training, confirm-ing a significant difference between imagery andverbal groups, mean change = +1.28 (SD = 5.86) vs.−6.69 (SD = 8.25), t(24) = 2.85, p = .0005 (onetailed), d = 1.13. However, because this differenceseemed to reflect mainly a decrease in the verbalgroup, we again tested for change in each groupseparately. Paired-sample t-tests indicated thatwithin the verbal condition emotionality ratingsdecreased significantly (i.e., became less positive)over time, t(12) = 2.93, p = .013, d = 1.10, butcontrary to expectation, there was no significantincrease in scores in the imagery condition,t(12) = 0.80, p = .44 (both tests two-tailed).

impact predictions about eachcondition

In order to check whether it was likely that thepositive affect results could be accounted for bydemand, participants were asked after the exper-iment to predict the impact of each separateexperimental condition on mood (yielding twoimpact predictions: one about the imagery andone about the verbal task condition). If partici-pants’ predictions about each training conditionwere in line with the actual results, then it couldbe argued that demand may have driven thereported differences in positive affect. Potential

244 holmes et al .

demand effects were investigated using a mixedmodel ANOVA in which the group factor wascondition (imagery versus verbal) and the within-subjects factor was condition impact prediction(prediction about imagery training versus predic-tion about verbal training). There was a maineffect of condition impact prediction, F(1,24) = 6.29, p = .019, η2 = .21. That is, overall,participants predicted that imagery training wouldhave a greater effect on emotion, relative tolistening “as normal” (mean rating = 3.50,SD = 3.77) than would the verbal condition(mean = 0.88, SD = 3.09). There was, however,no significant interaction between condition im-pact prediction, and participants' actual trainingcondition during the experiment, F(1, 24) = 0.27,p = .61; see Table 1.The association between the condition impact

predictions and actual changes in positive affectwas investigated using correlations. Within eachtraining condition, there were no significant corre-lations between anxiety change and impact predic-tions, although within the imagery condition thepredicted impact of the imagery task showed anonsignificant trend, r(11) = 0.51, p = .075. Oth-erwise the largest correlation, r(11) = −.45, p = .12,was for the predicted impact of the imagery taskwithin the verbal condition.Low sample sizes clearly limit the power of these

tests, so that all that can be concluded is that thecorrelations between individual participants’ pre-dictions about the possible impact of either trainingcondition (impact predictions) and the observedemotional effects are probably not very large. Whilewe found no significant association between actualanxiety change and predicted effects, the trend inthe imagery condition indicates that the possibilitythat demand may have had some influence cannotbe excluded.

subjective experience of listening totraining paragraphs

At the end of the experiment, participants wereasked questions about their subjective experienceof listening to the training paragraphs. The ratingsgiven in the imagery and verbal conditions werecompared using independent samples t-tests. Therewere several significant differences between con-ditions: participants in the verbal conditionreported using more verbal processing of thedescriptions during the training phase, ImageryM = 3.62 (SD = 2.32), Verbal M = 6.12(SD = 1.36), t(24) = 3.34, p = .003, d = 0.50;and participants in the imagery condition reportedusing more imagery during the training phase,Imagery M = 7.77 (SD = 1.42), Verbal M = 4.77

(SD = 2.24), t(24) = 4.07, p < .001, d = 1.60.These results indicate that participants reportedthat they were adhering to the appropriatecondition instructions during the training phase,although this may reflect demand characteristicsfollowing the experimental instructions.There was also some indication that at the test

phase of hearing the ambiguous paragraphs post-training, participants spontaneously deployed asimilar style of processing to the type they hadreceived during training. That is, those in the verbalcondition reported that they were verbally analyz-ing the meaning of the test sentences more thanparticipants in the imagery condition, ImageryM = 3.58 (SD = 1.83), Verbal M = 5.62(SD = 2.09), t(24) = 2.72, p = .012, d = 1.04.Conversely, those participants in the imagery condi-tion were more likely to report that they werethinking in images than those in the verbal condition,Imagery M = 7.33 (SD = 1.37), Verbal M = 4.77(SD = 2.09), t(24) = 3.59, p = .002, d = 1.45. Notethat no instructions had been given as to the style ofprocessing to be used during the test phase.

Discussionsummary of main results

The current experiment used overtly positivetraining material to investigate the effects ofimagery versus verbal interpretation training, incontrast to the use of benign material inExperiment 2 of Holmes and Mathews (2005).The present experiment also included a measureof positive affect to complement the measure ofstate anxiety used previously. The current resultsextended those previous findings and showedthat participants in an imagery training conditionreported greater increases in positive affect andgreater decreases in state anxiety after positiveinterpretation training than did those in a verbaltraining condition. We also found an absoluteincrease in positive affect (and a trend for stateanxiety to decrease) following imagery training.It is likely that either the improved positivetraining material used here promoted a morepowerful training effect or the measure ofpositive affect used here was more sensitive totraining effects, compared with the material andmeasures used by Holmes and Mathews (2005;Experiment 2). We suggest that these findingsprovide initial support for the hypothesis thatpositive imagery can have greater effects onpositive mood than does verbal processing ofthe same material. Needless to say, these resultsneed to be replicated in light of the small samplesize in the current study.

245mental imagery and pos i t i v e interpretat ion tra in ing

In addition to the predicted findings regardingmood change following training, there was also areduction in positive affect (and a trend for anincrease in state anxiety) from pre- to posttrainingin the verbal condition that had not been predicted.It may be that the task of doing the training trialswas arduous, and that without the counteractiveeffect of positive training with imagery, this causedthe relative decrease in positive affect. It is alsopossible that the verbal instructions and the focuson comprehending the paragraphs drew partici-pants’ attention to negative aspects of the trainingmaterial. Half of the paragraphs were constructedto begin with potentially negative scenarios thatwere then disambiguated in a benign manner(negative to benign). Thus, participants in theverbal condition may have focused more on thenegative components of the material. Futureresearch could test this possibility by using onlythe “benign to positive” training material. Howev-er, it is also possible that even some of the “benignto positive” material implied a degree of ambiguitythat could have suggested a potential negativeoutcome. For this reason, in future research itwould be of interest to include training materialdesigned to be completely unambiguous (“positiveto positive” scenarios).The results for the positively resolvable ambigu-

ous test descriptions (a measure of interpretativebias administered after the 10-min filler) suggestedthat emotionality ratings decreased significantly inthe verbal condition (i.e., became more negativelyvalenced) over time. Further, contrary to what onemight expect, there was no significant increase inemotionality scores in the imagery condition,though the results were in the expected direction(i.e., more positively valenced). No significantdifferences were found for the negatively resolvableambiguous test descriptions. Because there was nosignificant difference in mood measures betweenconditions after the 10-min filler, it is unlikely thatpersistent mood effects are sufficient to account forthe difference in the emotionality ratings betweenconditions.The finding of differences for the positively

resolvable rather than negatively resolvable scenar-ios may be explained by the fact that training usedspecifically positive materials. However, it is alsopossible that the interaction found for ratings of thepositively resolvable scenarios may depend in parton inflated ratings at pretraining. Further, theambiguous test descriptions used in the currentexperiment may have provided a less powerfulindex of interpretative bias than other measures ofinterpretative bias used in previous experiments,such as the recognition test used in Mathews and

Mackintosh (2000) or the lexical decision task usedin Hirsch, Clark, Williams, Morrison, and Math-ews (2005). Finally, it could be argued that themethod employed in this experiment and byHolmes and Mathews (2005) is unlike the “train-ing” used in other related studies. We used auditorypresentation of complete scenarios rather thanvisual presentation of partial sentences that weredisambiguated only by the final word. Althoughsome transfer effects to ambiguous test items werefound here, it could be argued the current methodrepresents a mood induction technique as much asit does interpretive training. Regardless of this issue,the current method has at least demonstrated theimportant role of imagery in producing positiveemotional change.Similar transfer of training effects to new test

paragraphs were found by Holmes and Mathews(2005) in a first experiment where the test para-graphswere given immediately posttraining, but notin a second experiment in which a verbal filler task(answering questions about a do-it-yourself homeimprovement manual) was interpolated betweentraining and test. Because we had thought that theverbal nature of this filler task may have disruptedtransfer of the trained processing style in theimagery condition, in the present experiment anonverbal filler task was used. The current findingswith this new filler task lend some support to thisidea. However, since several variables were changedbetween the experiments (e.g., training material andmeasures as well as the filler task), strong conclu-sions cannot be drawn at this stage.

overall conclusions andimplications for future research

To date, much of the literature on the interpretationtraining paradigm has focused on the impact ofnegative training. In contrast, the positive effects ofinterpretation training have, as yet, received lessresearch attention. The current experiment istherefore an important early step indicating thepotential for positive training effects on mood.However, a major limitation of the current study isthe small sample size used, which means that anyconclusions drawn must remain tentative untilreplicated. The current experiment has howeverprovided a new method to test the differentialimpact of imagery versus verbal instructions forpositive interpretation training. New positive train-ing material was developed, and a measure ofpositive affect included. Overall, we conclude thatthe results are consistent with the hypothesis thatimagery may have a more powerful impact onemotional responses than verbal processing of thesame material, and there is now initial support that

246 holmes et al .

this assumption applies to positive as well asnegative affect (cf. Holmes & Mathews, 2005).These findings indicate that emotional effects ofimagery may be agnostic with regards to valence,rather than being valence specific. Future researchshould seek to replicate the results of this positivetraining experiment in order to establish thereliability of the findings. From a theoreticalperspective, use of the interpretation trainingparadigm provides only one method by which totest the overarching hypothesis that imagery mayhave a more powerful impact on emotionalresponse than verbal processing of the samematerial. It is important to also seek convergentevidence for the current findings using alternativemethodological paradigms.

Clinical ImplicationsThe current findings, in conjunction with those ofHolmes and Mathews (2005) provide an empiricaljustification for the use of imagery as a device toevoke—and perhaps also to modify—emotionalresponses in clinical conditions. From a clinicalperspective, a key issue for further investigation iswhether reductions in negative affect may best bepromoted using imagery rather than verbal proces-sing in clinical populations, and what types ofimagery techniques could be used. If imagery has aspecial relation with emotion, then this suggests thata variety of therapeutic approaches might be usefulin reducing the impact of negative imagery. Giventhat patients with clinical disorders have a negativebias that appears to cause vulnerability to anxiety(Mathews&MacLeod, 2002), then clinically it is ofgreat interest to develop effective techniques topromote positive bias, such as computerized inter-pretation training. Imagery instructions may pro-vide a useful tool by which to enhance such trainingtechniques. Positive imagery training may help therelatively automatized production of benign orpositive imagery when encountering novel stimuli.It would also be interesting to test whether similartraining could be used to reduce negative biasassociated with emotional disorders other thananxiety, such as major depressive disorder. Indeed,this special issue reflects the current zeitgeist forclinical applications of experimental training para-digms. Further work examining the training ofpositive mood and positive biases in clinicalpopulations is now needed.A potentially important implication for the

interpretation training literature is that typicaltraining effects found on affect measures andmeasures of emotional bias (emotionality ratings)were not achieved with verbal instructions alone.

This finding underscores the importance of usingimagery instructions, as these seem likely to be anactive rather than incidental component of theoriginal procedure developed by Mathews andMackintosh (2000) in creating affect change.Further, whatever the explanation, the initialfinding that verbal instructions led to a reductionin positive mood after exposure to positive infor-mation, and to more negative resolutions ofambiguous test material, may have interestingclinical implications. It is possible that in certainsituations, promoting verbal positive thoughts incognitive therapy might not only have less impactthan imagery in promoting positive mood, but evenlead to a reduction in positive mood. Clearly,however, further research is needed before extrap-olating such conclusions to a clinical setting.A methodological step that could be useful for

future studies is the instructions developed toincrease participants’ awareness of using mentalimagery through the “imagine cutting a lemon”practice. Several participants were unsure of whatwas meant by “using mental imagery” and theexamples in the instructions provide a usefulsubjective illustration to add clarification and useof imagery. Holmes and Mathews (2005), alongwith the current experiment and Mackintosh,Mathews, Yiend, Ridgeway, and Cook (2006; thisissue), have been the first to deliver trainingscenarios in an auditory verbal format, ratherthan a written text-based format. Postexperimentaldebriefing indicated that, despite the lengthyexperimental session (approximately 2 hours), theparticipants maintained interest and concentrationwhile listening to the stimuli. This observationappears to contrast with previous interpretationtraining studies (e.g., Mathews & Mackintosh,2000) in which participants have reported thatreading the volume of training material on thecomputer screen can be rather tedious. Auditorypresentations of text-based material may thereforebe useful in developing future clinical applicationsand might also provide a more portable format forhome use.Informal observations indicate that while some

clinicians have made sweeping claims about imag-ery in therapy, others—perhaps deterred by the lackof evidence—have focused mainly on verbalthoughts rather than images. However, we notethat in Beck’s original conception of cognitivetherapy (Beck, 1976), the importance of assessingpatients’ images, as well as their verbal thoughts,was emphasized. Our impression is that thisemphasis seems to have been neglected in muchrecent clinical practice and training, at least inEurope. For example, cognitive therapy techniques

247mental imagery and pos i t i v e interpretat ion tra in ing

are typically taught through the use of verballyevaluating negative automatic thoughts and bykeeping daily written diaries. This strategy canperhaps lead therapists and clients to focus onverbal thoughts and neglect the impact of imagery.A dominant focus on verbal thoughts can leadtherapists and clients to miss other cognitiveprocesses that may be causal in the maintenanceof a given psychological disorder and perhaps keyto successful treatment (e.g., Hirsch, Clark, Math-ews, & Williams, 2003; Holmes & Hackmann,2004). The current evidence for a special impact ofimagery on emotion may serve to encouragetherapists to consider using imagery techniquesnot just in those psychological disorders where theiruse is standard (such as posttraumatic stressdisorder; e.g., Brewin & Holmes, 2003) but toconsider using imagery in disorders where it hasbeen less well explored (e.g., depression).

ReferencesBaddeley, A. D., & Andrade, J. (2000). Working memory and

the vividness of imagery. Journal of Experimental Psychol-ogy–General, 129, 126–145.

Beck, A. T. (1976). Cognitive therapy and the emotionaldisorders. New York: International Universities Press.

Brewin, C. R., & Holmes, E. A. (2003). Psychological theoriesof posttraumatic stress disorder. Clinical Psychology Re-view, 23, 339–376.

Grey, S., & Mathews, A. (2000). Effects of training oninterpretation of emotional ambiguity. Quarterly Journalof Experimental Psychology Section A—Human Experi-mental Psychology, 53, 1143–1162.

Hertel, P. (2002). Cognitive biases in anxiety and depression:Introduction to the special issue. Cognition and Emotion,16, 321–330.

Hirsch, C. R., Clark, D. M., Mathews, A., & Williams, R.(2003). Self-images play a causal role in social phobia.Behaviour Research and Therapy, 41, 909–921.

Hirsch, C. R., Clark, D. M., Williams, R., Morrison, J., &Mathews, A. (2005). Interview anxiety: Taking the perspec-tive of a confident other changes inferential processing.Behavioural and Cognitive Psychotherapy, 33, 1–12.

Holmes, E. A., and Hackmann, A. (Eds.). (2004). Mentalimagery and memory in psychopathology [Special issue].Memory, 12(4).

Holmes, E. A., & Mathews, A. (2005). Mental imagery andemotion: A special relationship? Emotion, 5, 489–497.

Mackintosh, B., Mathews, A., Yiend, J., Ridgeway, V., & Cook,E. (2006). Induced biases in emotional interpretation trainingendure despite changes in context. Behavior Therapy, 37, xx.

Marks, D. F. (1973). Visual imagery differences in the recall ofpictures. British Journal of Psychology, 64, 17–24.

Mathews, A., & Mackintosh, B. (2000). Induced emotionalinterpretation bias and anxiety. Journal of AbnormalPsychology, 109, 602–615.

Mathews, A., & MacLeod, C. (2002). Induced processingbiases have causal effects on anxiety. Cognition andEmotion, 16, 331–354.

Reisberg, D., Pearson, D. G., & Kosslyn, S. M. (2003).Intuitions and introspections about imagery: The role ofimagery experience in shaping an investigator's theoreticalviews. Applied Cognitive Psychology, 17, 147–160.

Spielberger, C. D., Gorsuch, R. L., Lushene, R., Vagg, P. R., &Jacobs, G. A. (1983). Manual for State-Trait AnxietyInventory. Palo Alto, CA: Consulting Psychologists Press.

Watson, D., & Clark, L. A. (1994). The PANAS-X—Manual forthe Positive and Negative Affect Schedule—Expanded form.Unpublished manuscript, University of Iowa, Iowa City.

Watson, D., Clark, L. A., & Tellegen, A. (1988). Developmentand validation of a brief measure of positive and negativeaffect. Journal of Personality and Social Psychology, 54,1063–1070.

RECEIVED: August 9, 2005ACCEPTED: February 9, 2006Available online 24 May 2006