Applied Animal Behaviour Science 118 (2009) 161–181

Cognitive bias as an indicator of animal emotion and welfare: Emergingevidence and underlying mechanisms§

Michael Mendl *, Oliver H.P. Burman, Richard M.A. Parker, Elizabeth S. Paul

Centre for Behavioural Biology, Department of Clinical Veterinary Science, University of Bristol, Langford House, Langford, BS40 5DU, UK

A R T I C L E I N F O

Article history:

Available online 21 March 2009

Keywords:

Emotion

Cognitive bias

Animal welfare

Attention

Reward valuation

Subjective probability

A B S T R A C T

Accurate assessment of animal emotion (affect) is an important goal in animal welfare

science, and in areas such as neuroscience and psychopharmacology. Direct measures of

conscious emotion are not available, so assessment of animal affect has relied on measures

of the behavioural and physiological components of affective states. These are important

indicators but have some limitations (e.g. measuring emotional arousal rather than

valence (positivity vs negativity)). Human psychology research suggests that changes in

cognitive function (information processing) can also be reliable indicators of emotional

state (especially valence). For example, people in negative states attend to threats, retrieve

negative memories, and make negative judgements about ambiguous stimuli more than

happier people. Here we review a new research area investigating the possibility that such

affect-induced ‘cognitive biases’ also occur in animals. We focus on a novel ‘judgement

bias’ paradigm in which animals are trained that one cue predicts a positive event and

another cue predicts a less positive/negative event, and are then presented with

ambiguous (intermediate) cues. The hypothesis is that animals in a negative affective state

will be more likely to respond to (‘judge’) these ambiguous cues as if they predict the

negative event (a ‘pessimistic’ response), than animals in a more positive state. Recent

studies of rats, dogs, rhesus monkeys, starlings and humans provide face-value support for

this hypothesis. We discuss the strengths and weaknesses of the affect manipulation

treatments used in these studies, and the possibility that treatment-induced changes in

feeding motivation, general activity and learning are responsible for the effects observed,

and we consider whether the type of bias observed and the precise design of the

judgement bias task can provide information about different types of affective state.

Judgement biases may result from the influence of affect on decision-making processes

including attention to and perception of the ambiguous cue, evaluation of the value and

probability (expected utility) of the outcomes of different responses, and action selection.

Affect might also modulate general tendencies of loss, risk and ambiguity aversion, hence

biasing decisions. We discuss these possibilities in relation to theory and findings from

neurobiological and psychological studies of decision-making, in order to better

understand the potential mechanisms underlying judgement biases. We conclude with

some specific recommendations for study design and interpretation, and suggestions for

future research in this area.

� 2009 Elsevier B.V. All rights reserved.

Contents lists available at ScienceDirect

Applied Animal Behaviour Science

journa l homepage: www.e lsev ier .com/ locate /applan im

§ This paper is part of a special issue entitled ‘‘Animal Suffering and

Welfare’’, Guest Edited by Hanno Wurbel* Corresponding author. Tel.: +44 117 928 9485; fax: +44 117 928 9582.

E-mail address: mike.mendl@bris.ac.uk (M. Mendl).

0168-1591/$ – see front matter � 2009 Elsevier B.V. All rights reserved.

doi:10.1016/j.applanim.2009.02.023

1. Introduction

Concerns for animal welfare are generally based on theassumption that non-human animals can subjectivelyexperience emotional (affective) states and hence can

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181162

suffer or experience pleasure (Dawkins, 1990; Mendl,2001; Mendl and Paul, 2004; Boissy et al., 2007). Recentpolitical and legal statements echo the basis of theseconcerns. For example, European legislation aims ‘toensure improved protection and respect for the welfareof animals as sentient beings’ (italics added; EuropeanUnion, 1997. Treaty of Amsterdam, Protocol on Protectionand Welfare of Animals (p. 110)), and the Australiangovernment’s strategy for animal welfare covers the careand use of ‘all sentient species in Australia’, where asentient animal is defined as ‘one that has the capacity tohave feelings and to experience suffering and pleasure’(Australian Government, 2008. Australian Animal WelfareStrategy (p. 7)). Although direct measurement of sub-jective emotional experiences is not currently possible (fora different view, see Wemelsfelder, 1997), the develop-ment of accurate proxy measures is therefore an importantgoal in animal welfare science, as well as in otherdisciplines such as neuroscience and psychopharmacology(Panksepp, 1998; Mendl and Paul, 2004; Rolls, 2005;Lawrence, 2008).

Most emotion researchers consider that emotions arisein situations that are ‘important’ to the organism, in thesense that they may influence its survival and reproductivesuccess. The primary function of emotions in thesecontexts is widely hypothesised to be to guide the animal’sbehavioural decisions in order to achieve survival goals –the attainment of valuable resources/rewards, and theavoidance of harm/punishment – perhaps by providing a‘common currency’ that the animal uses to determinewhich behaviour or sequence of behaviours is most likelyto enhance survival (e.g. Ortony et al., 1988; Cabanac,1992; Oatley and Jenkins, 1996; Cardinal et al., 2002; Rolls,2005).

In order to study emotional states scientifically, thedevelopment of accurate measures is an essential first step.The measurement of subjective experience in non-humanspecies is fraught with difficulty, and debates continue asto whether and which non-human animals (hereafter‘animals’) have the capacity for such experience (e.g.Carruthers, 1989; Kennedy, 1992; Griffin, 1992, 1998;Macphail, 1998; Baars, 2001; Bermond, 2001; Panksepp,2005). However, emotional states are recognised byhuman psychologists as being multifaceted and compris-ing other ‘components’ in addition to subjective experi-ence, namely behavioural and physiological changes (e.g.Plutchik, 1980; Ekman, 1984; Scherer, 1984; Frijda, 1988;Smith and Lazarus, 1993; Clore and Ortony, 2000; Lernerand Keltner, 2000). For example, the emotion of fearincludes the subjective experience of fear, but also theexpression of freezing or fleeing behaviour, and alterationsin physiology such as changes in heart rate, blood pressure,and circulating glucocorticoids. In humans, linguisticreport can be used as a measure of a person’s subjectiveemotional experience that is likely to be as reliable anindicator as any (though not infallible), but this is clearlynot possible in animals. Instead, behavioural and physio-logical indicators form the basis for nearly all currentindicators of animal emotional states (e.g. approach/avoidance behaviour; vocalizations; play behaviour;behavioural tests such as open field, elevated plus maze

(EPM), light–dark box test, sucrose consumption, forcedswim; indicators of hypothalamic–pituitary adrenal (HPA)and sympathetic–adrenomedullary (SAM) activity; otherendocrine indicators such as oxytocin; see Paul et al.,2005).

These indicators offer a great deal of information.However, they are not free from problems of interpretation(Paul et al., 2005). Some may be good measures ofemotional arousal (intensity, or how ‘activated’ the animalis), but less good measures of emotional valence (whetherthe emotional state is positive or negative; see Watsonet al., 1988; Russell, 2003). For example, HPA and SAMactivity may increase in a range of situations includingthose that are likely to have quite different emotionalvalence (e.g. meeting a predator vs meeting a sexualpartner) or which may be affectively neutral (e.g. increasedlocomotor activity) (Rushen, 1986, 1991; Baldock et al.,1988; Marchant et al., 1995). From an animal welfareperspective these are significant problems because,although the arousal and intensity of emotional statesare important to know about, whether these states arepositive or negative for the animal (valence) is the criticalmeasure.

A related issue is that many measures lack a priori

hypotheses for how they should change according to theanimal’s emotional state (specifically, its valence). This canmake interpretation of tests, and translating them fromone species to another, difficult. For example, does adecreased latency to stop swimming in the forced swimtest reflect a state of depression or despair, or does itrepresent an adaptive coping response (Cryan andMombereau, 2004)? What are the predictions for testsdesigned to measure anxiety-related ‘wall-hugging’ thig-motaxis (e.g. open field test, EPM) or preference for darkareas (e.g. EPM, light–dark box test) when adapted for usein a diurnal species with low fear of light or open spaces (cf.Janczak et al., 2002)?

Other limitations include few measures of positive

affective states, despite these being of increasing interestin animal welfare research (Boissy et al., 2007; Yeates andMain, 2008), and the finding that, in humans, linguisticreport of subjective emotion may dissociate from otherindicators of emotional state. For example, some peoplereport no change in subjective emotional experience whileexhibiting clear physiological indicators of an emotionalresponse, while others report emotional experiences in theabsence of the expected physiological changes (e.g. Patricket al., 1993; Stone and Nielson, 2001). While suchdissociations provide valuable information about therelationships between components of emotion, they alsoraise uncertainty as to the extent to which existingbehavioural and physiological indicators map on to thesubjective emotional states which lie at the heart of animalwelfare concerns.

Given the above issues (described in more detail by Paulet al., 2005), there is clearly room for the development ofnew methods for assessing animal emotion. One majorgrowth area is the study of neural correlates of emotionalresponses both in humans (primarily using functionalmagnetic resonance imaging) and animals (using a rangeof techniques including single-cell recording, lesioning

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 163

studies and microdialysis). Many brain areas, including theamygdala, prefrontal and orbitofrontal cortex (OFC),anterior cingulate, insula, nucleus accumbens (NAc),ventral tegmental area, and periaqueductal grey, and brainsystems such as the dopamine, noradrenaline, opioid andserotonin systems, appear to be involved in the processingof affective information (e.g. Bechara et al., 1994; LeDoux,1996; Panksepp, 1998; Damasio, 2000; Lang et al., 2000;Spruijt et al., 2001; Cardinal et al., 2002; Berridge, 2003;Camille et al., 2004; Rolls, 2005; Coricelli et al., 2007;Mobbs et al., 2007; Murray, 2007; Leknes and Tracey,2008), including the encoding of reward value andprobability that appear to be critical determinants ofdecision-making (e.g. Knutson et al., 2005; Padoa-Schioppa and Assad, 2006; Paton et al., 2006; Rangelet al., 2008). Research in this area is revealing the neuralsubstrates and mechanisms underlying emotional pro-cesses and their role in guiding animal decisions. However,measures of neural function are technically complex anddo not currently offer a practical means for assessingemotional states in unrestrained animals.

One approach which does offer some promise concernsthe fourth, cognitive, component of emotion. Alongsidesubjective, behavioural and physiological components, ithas long been recognised in human psychology thatcognitive processes influence and are influenced by anindividual’s emotional state. We follow Shettleworth(1998) in using the term ‘cognitive’ broadly to refer toinformation processing including attention, learning,memory and decision-making. The role of cognitiveprocesses in influencing the emotional response that anindividual makes to an event or situation is thought tooccur through a process of ‘appraisal’ (Schachter andSinger, 1962) wherein the result of a series of stimuluschecks (e.g. is the stimulus sudden, familiar, predictable,pleasant?) determines the emotional response that ismade. These checks are ‘cognitive’ in the sense that theyinvolve some evaluation of the information presented byan event and may, for example, involve reference tomemories of previous events, but they may also berelatively simple, rapid and automatic (e.g. Grandjeanand Scherer, 2008). A number of appraisal theories exist(e.g. Oatley and Johnson-Laird, 1987; Lazarus, 1991;Scherer, 2001), and Scherer’s theory has been used byBoissy and colleagues as a basis for studying animalemotions. By presenting animals with stimuli that havespecific characteristics (e.g. sudden, unfamiliar, unpredict-able, unpleasant) known to map on to reported emotionalstates in people (e.g. fear), the animals’ resulting profile ofbehavioural and physiological responses may then be usedas an indicator of those corresponding emotional states(e.g. Desire et al., 2002, 2004, 2006). If one accepts theassumption that the relationships between stimulusappraisals and emotions found in people hold true forother species, this paradigm offers a useful a priori

approach to identifying indicators of putative emotionalstates.

While cognitive processes can influence emotions,emotional state can also influence cognition. This opensup the possibility of using clearly defined, objectivemeasures of cognitive performance as indicators of the

more elusive and less easily measured emotional statesthat influence them. It is this aspect of the cognition–emotion relationship that we will focus on.

2. The influence of emotion on cognition: ‘cognitive bias’

Studies of human psychology provide the main sourceof information about the effects of emotional states oncognitive function. As with the work on appraisal, there is arisk that such effects do not hold for other species.However, if we are ultimately interested in the subjectiveemotional experiences of animals, perhaps the only ‘modelsystem’ we can turn to is the human being who canlinguistically report on subjective emotion, and thusprovide information on how this component of emotionis related to others that we can measure in animals(behaviour, physiology, cognitive function). In fact, con-temporary researchers, including animal welfare scien-tists, are increasingly seeking parallels between humansand animals in their studies of emotion, cognition,consciousness and mind (e.g. Clark and Squire, 1998; Calland Carpenter, 2001; Desire et al., 2002; Garner et al.,2003; Smith et al., 2003; Paul et al., 2005; Herry et al.,2007; Shafir et al., 2008).

A consistent finding from human studies is that thevalence of an individual’s emotional state appears toinfluence a number of cognitive processes includingattention, memory and judgement. We review thesefindings in detail in Paul et al. (2005), and only summarisethe main points here. Many studies, often using compu-terised tasks, have demonstrated that people reporting anegatively valenced affective state (e.g. anxiety) showenhanced attention to threatening stimuli, such as imagesof angry faces, snakes or negative words, relative to peoplein a more positive state (e.g. Mathews and Macleod, 1994;Mineka et al., 1998; Mogg and Bradley, 1998). Likewise,there is evidence that affective state influences memoryretrieval, with happier people being more likely to recallpositive memories and unhappy or depressed people morelikely to recall negative ones (e.g. Bower, 1981; Burke andMathews, 1992; Denny and Hunt, 1992; Mineka et al.,1998). Furthermore, people in a negative state are morelikely to make negative judgements about future events orambiguous stimuli (‘pessimism’) than people in positivestates who show more optimistic judgements and inter-pretations (e.g. Eysenck et al., 1991; Wright and Bower,1992; MacLeod and Byrne, 1996; Nygren et al., 1996).

There are inevitably some uncertainties in theseresearch findings. For example, it is not always the casethat similarly valenced states, such as depression, anxiety,fear and anger, have the same effects on cognitive function(Lerner and Keltner, 2000). Additionally, some effects maybe the result of trait, as well as or instead of, statedifferences in affect (Mineka et al., 1998; Mogg andBradley, 2005). Nevertheless, taken together, the findingssuggest that measures of attention, memory and judge-ment biases can often be indicative of a person’s emotionalstate. If this is also the case in animals, it may offer anumber of advantages to behavioural and physiologicalindicators of emotion. These include specifically measur-ing emotional valence; providing general a priori predic-

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181164

tions, based on the findings of human studies, for howcognitive performance and emotion co-vary and henceallowing tests to be readily applied across differentspecies; and allowing the measurement of positiveaffective states. Furthermore, to date it appears that biasesin information processing closely reflect, and thereforemay be reliable indicators of, reported subjective emo-tional experience in humans, and hence possibly inanimals too. However, active investigation of dissociationsbetween subjective emotion and cognitive performancehas not been a major focus of previous research, and suchdissociations may be found.

The modulating effects of emotional valence oncognitive function are hypothesised to have adaptivevalue (e.g. Mineka et al., 1998; see also Haselton andNettle, 2006). For example, an individual in an anxiety orfear-inducing situation would benefit from directingattention to threatening stimuli, storing and retrievingmemories of what happens in such situations, and making‘safety first’ judgements about ambiguous stimuli, such asa rustle in the grass, that assume the worst. If so, selectionwould favour the evolution of these modulating effects inother species as well as humans, hence making the searchfor cognitive indicators of emotional state in animals aplausible endeavour.

2.1. ‘Cognitive bias’ terminology

In the human literature, the term ‘cognitive bias’ hasbeen used as a general label for the effects of emotionalstate or trait on cognitive processes (e.g. Mineka andSutton, 1992; Mathews et al., 1995; Warda and Bryant,1998). This term is now also used to describe newmeasures of cognitive function in animals that are beingused as proxy indicators of affective state (e.g. Hardinget al., 2004; Paul et al., 2005; Bateson and Matheson, 2007).In this context, the term is a useful shorthand to describethe area of study – the influence of affect on informationprocessing in animals (strictly we might call it ‘affect-induced cognitive bias’) – but should not be taken to implythat the processes underlying the effects seen arecognitively complex or well understood (see below), orthat they involve either subjective experience of emotion,or conscious thought processes. Indeed, there is consider-able evidence that affective states can influence pre-conscious attentional processes in people (e.g. Mathewsand Macleod, 1994; Bradley et al., 1995). Furthermore, were-emphasise that while cognitive bias measures mayprovide reliable indicators of the valence of emotionalstates, they, like ‘classical’ behavioural and physiologicalindicators, cannot demonstrate unequivocally whethersuch states are consciously experienced in animals. Thisremains a separate and, at present, largely philosophicaldiscussion which attracts widely differing views as notedabove.

In this paper, we use ‘cognitive bias’ as an umbrellaterm to cover the influence of affect on a range of cognitiveprocesses including attention, memory and judgement.We focus on studies of bias in the judgement of ambiguousstimuli as these have been the main area of research inanimals to date. While we refer to these as ‘judgement

biases’, this should not be taken to indicate that a distinct‘judgement’ mechanism underlies such biases. As willbecome clear later, it is quite possible that changes inattention, memory and other processes, such as evaluationof the value and probability of different decision outcomes,may be responsible for the biased judgements revealed byour experimental paradigm.

A final theoretical point is that although the term ‘bias’may suggest some irrational decision-making process, anapparently biased decision may make perfect adaptivesense if everything is known about the animal’s state,knowledge of the situation, and past experience, and hencethe prior information on which its decision is based. Giventhis ‘prior’ information, Bayesian statistical approaches canbe used to model the expected decision (cf. Trimmer et al.,2008).

3. Cognitive bias as an indicator of emotionalstate in animals

3.1. Judgement bias: the Harding et al. (2004) cognitive

bias study

The majority of studies so far carried out on cognitive biasin animals have studied ‘judgement biases’. Judgement biasin this context refers to the propensity of a subject to showbehaviour indicating anticipation of either relativelypositive or relatively negative outcomes in response toaffectively ambiguous stimuli. Such propensities can beoperationally defined as ‘optimism’ or ‘pessimism’ respec-tively (see Matheson et al., 2008) with the caveat that thisshould not be taken to imply any conscious equivalent ofthese states as experienced by humans. As mentionedabove, biases in judgement may result from affectiveinfluences on a variety of cognitive processes.

The first published study on cognitive bias in animals(Harding et al., 2004) introduced the judgement biasexperimental paradigm that has formed the basis for mostsubsequent studies (Fig. 1). In this operant discriminationtask, rats were trained to perform one response (P) whenpresented with one cue (p: a tone of a particular frequency(e.g. 2 or 4 kHz)) in order to experience a positive event(e.g. arrival of a food pellet), and to perform a differentresponse (N) when presented with a different cue (n: atone of a different frequency (e.g. 4 or 2 kHz)) in order toavoid a negative event (e.g. a burst of white noise). Fortechnical reasons the responses used in this study werelever press (P) and no lever press (N)—a ‘go/no-go’ task.Once trained on this discrimination task, rats were thentested by presenting them with occasional unreinforcedambiguous ‘probe’ cues (tones of intermediate frequencybetween p and n). Harding et al. (2004) used three probecues, each presented with a probability of 0.085 inter-spersed between standard training trials on the twotrained cues. Following the findings of human studies(see earlier), we hypothesised that animals in a negativeemotional state would be more likely to categorise theambiguous cues as predicting the negative event andhence would be more likely to show response N.

To induce presumed differences in affective state,Harding et al. (2004) housed rats that had been trained

Fig. 1. Experimental paradigm for the original cognitive (judgement) bias task (Harding et al., 2004). See text for explanation.

Fig. 2. Proportion of cues responded to with a lever press (a), and latency

to lever press (b) during test sessions in which rats were presented with

either training cues (‘food tone’ = tone predicting positive event (food);

‘noise tone’ = tone predicting negative event (white noise)) or ambiguous

probe cues. During this phase of the study, subjects were kept either in

unpredictable (filled circles) or predictable (open circles) housing

conditions (see text for details). Redrawn with permission from

Harding et al. (2004) Nature 427, 312.

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 165

on the above task, in either unpredictable housingconditions similar to those used in previous work toinduce a mild stress/depression-like state (Willner, 1997;Zurita et al., 2000; Porsolt, 2000), or in predictable, controlconditions. When tested under these housing treatments,the rats showed responses to probe cues in the judgementbias task that were consistent with our predictions (Fig. 2).Specifically, rats in unpredictable housing showed atendency to make a lower proportion of positive responses(lever presses) to the probe tone closest to the food toneand to the food tone itself (Fig. 2a), and were significantlyslower in making these responses to these tones comparedto rats in the control conditions (Fig. 2b). The findingsindicated that rats in a putative negative state were lesslikely to respond to probe tones as if they indicated apositive event than were control animals, similar to thediminished anticipation of positive events (decreasedoptimism) observed in depressed humans (e.g. Wrightand Bower, 1992; MacLeod and Byrne, 1996).

These initial results were very encouraging, but therewere some unexpected findings. For example, treatmentgroups differed in their response to one of the trained cues(the food tone), whereas one might predict that responsesto the unambiguously reinforced cues should remainunaffected, with treatment differences only being evidentfor the ambiguous probe cues. One possible explanation forthis is that the treatment may have resulted in affect-induced, or affect-independent, differences in: bodyweight and consequent feeding motivation and hencealtered response to the food tone; anhedonia whichdecreased the rewarding value of the sweet food pelletsused in the task (see Willner, 1997); or general activitywhich, due to the go/no-go nature of the task, might havecaused a drop off in lever pressing across all tones. Toinvestigate these possibilities, all rats were weighed andsubjected to independent tests of feeding motivation (timeto consume 50 of the food pellets used in the task, cf.Nielsen, 1999; Whishaw et al., 1992), anhedonia (pre-ference for and consumption of a sweet sucrose solution,Papp and Wieronska, 2000; Zurita et al., 2000), and activity(holeboard test, Fernandes and File, 1996). In addition,lever press responses to the two training tones immedi-ately before and after the housing treatment were analysedto check that the treatment groups did not differ in howwell they had learnt the task. No differences were found inany of these analyses.

Although it is not possible to provide definitiveexplanations for the findings of this study, Hardinget al.’s (2004) paper suggested that there was clearpotential for the use of judgement bias as an indicatorof animal emotion. Subsequently, a number of studies havebeen carried out in different species, and using modifica-tions of the original judgement bias task, to test thegenerality and robustness of the Harding et al. (2004)findings. These studies have also addressed specific issuessuch as how to speed up the discrimination learning at thecore of the task to make the approach more practicallyfeasible.

Ta

ble

1

Su

mm

ary

of

rece

nt

stu

die

so

fju

dg

em

en

tb

ias

as

an

ind

ica

tor

of

aff

ect

ive

sta

te.

Sp

eci

es

an

dst

ud

ya

Cu

ety

pe

Re

spo

nse

bP

osi

tiv

ere

info

rce

rN

eg

ati

ve

rein

forc

er

Aff

ect

ma

nip

ula

tio

nP

red

icte

db

ias?

c

Ra

t(1

)T

on

efr

eq

ue

ncy

Go

/no

-go

(le

ve

rp

ress

)Fo

od

No

ise

Un

pre

dic

tab

leh

ou

sin

gY

es:

pro

be

ne

ar

+v

ecu

e

Ra

t(2

)S

pa

tia

llo

cati

on

Go

/no

-go

(lo

com

oti

on

)Fo

od

No

foo

dE

nri

chm

en

tY

es:

pro

be

ne

ar�

ve

cue

Ra

t(3

)S

pa

tia

llo

cati

on

Go

/no

-go

(lo

com

oti

on

)Fo

od

Un

pa

lata

ble

foo

dB

rig

ht

lig

ht

vs

dim

lig

ht

Yes

:a

cro

ssa

llp

rob

es

Ra

t(4

)T

on

efr

eq

ue

ncy

Act

ive

cho

ice

(le

ve

rp

ress

)2

foo

dp

ell

ets

1fo

od

pe

lle

tU

np

red

icta

ble

ho

usi

ng

No

Ra

t(4

)T

on

efr

eq

ue

ncy

Act

ive

cho

ice

(le

ve

rp

ress

)2

foo

dp

ell

ets

1fo

od

pe

lle

tE

nri

chm

en

tN

o

Do

g(5

)S

pa

tia

llo

cati

on

Go

/no

-go

(lo

com

oti

on

)Fo

od

No

foo

dS

ep

ara

tio

na

nx

iety

sco

reY

es:

at

cen

tra

lp

rob

e

Hu

ma

n(6

)V

isu

al

(lo

cati

on

on

scre

en

)A

ctiv

ech

oic

e(k

ey

pre

ss)

Nic

eim

ag

ea

nd

ga

inp

oin

ts

Na

sty

ima

ge

an

d

lose

po

ints

Ind

ivid

ua

lv

ari

ati

on

inm

oo

dY

es:

at

cen

tra

l(o

nly

)p

rob

e

Hu

ma

n(6

)V

isu

al

(lo

cati

on

on

scre

en

)A

ctiv

ech

oic

e(k

ey

pre

ss)

Nic

eim

ag

ea

nd

ga

inp

oin

ts

Na

sty

ima

ge

an

d

lose

po

ints

Mu

sic

mo

od

ma

nip

ula

tio

nY

es:

at

cen

tra

l(o

nly

)p

rob

e

Rh

esu

sm

aca

qu

e(7

)V

isu

al

(lin

ele

ng

th)

Go

/no

-go

(to

uch

scre

en

)Fo

od

No

ise

Ve

teri

na

ryin

spe

ctio

nY

es:

at

cen

tra

lp

rob

ea

nd

pro

be

ne

ar

+v

ecu

e

Sta

rlin

g(8

)V

isu

al

(gre

ysc

ale

)G

o/n

o-g

o(l

id-fl

ipp

ing

)Fo

od

Un

pa

lata

ble

foo

dE

nri

chm

en

tY

es:

pro

be

ne

ar

+v

ecu

e

Sta

rlin

g(9

)V

isu

al

(ke

yil

lum

ina

tio

n

du

rati

on

)

Act

ive

cho

ice

(co

lou

red

ke

yp

eck

)

‘in

sta

nt’

foo

d

(aft

er

1s)

De

lay

ed

foo

d(a

fte

r1

5s)

En

rich

me

nt

Yes

:p

rob

en

ea

r�

ve

cue

aS

tud

yid

en

tifi

cati

on

:(1

)H

ard

ing

et

al.

(20

04

);(2

)B

urm

an

et

al.

(20

08

a);

(3)

Bu

rma

ne

ta

l.(s

ub

mit

ted

);(4

)P

ark

er

(20

08

);(5

)C

ase

ye

ta

l.(2

00

8);

(6)

Me

nd

le

ta

l.(2

00

6);

(7)

Be

the

lle

ta

l.(2

00

7);

(8)

Ba

teso

na

nd

Ma

the

son

(20

07

),(9

)M

ath

eso

ne

ta

l.(2

00

8).

b‘G

o/n

o-g

o’

stu

die

sy

ield

me

asu

res

of

bo

thth

eo

ccu

rre

nce

or

no

n-o

ccu

rre

nce

of

the

‘go

’re

spo

nse

,a

nd

the

late

ncy

tom

ak

eth

e‘g

o’

resp

on

se,

bo

tho

fw

hic

hca

nb

eu

sed

tote

stju

dg

em

en

tb

ias

pre

dic

tio

ns.

cP

red

icti

on

sa

reb

ase

do

nth

eh

yp

oth

ese

sfr

om

hu

ma

nst

ud

ies

sum

ma

rise

din

the

tex

t.

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181166

3.2. Judgement bias: studies of rats, humans, rhesus monkeys,

dogs and starlings based on the Harding et al. (2004)

paradigm

Table 1 summarises a number of recent studies ofjudgement bias in animals, including several completedbut as yet unpublished findings, most of which have beenpeer reviewed for conference presentation. The tableshould thus be taken as an indication of progress to date,with the caveat that full publication in peer-reviewedjournals awaits some studies. Two of our studies ofhumans are included in the table because they used tasksbased directly on the original Harding et al. (2004)paradigm. This allowed us to investigate whether previousfindings in humans, based largely on linguistic tests ofcognitive bias and providing the rationale for the animalcognitive bias work, would be reflected in an inherentlynon-linguistic task (judging the location of a cross on aline) that parallels tasks used in animals, hence offering theopportunity for direct comparison of the cognitive biasphenomenon between humans and other species.

We now briefly summarise key points of the studymethods (Table 2). Study 1 is the Harding et al. (2004)experiment. Study 7 used a very similar design with foodand noise reinforcers for rhesus monkeys, except that thetraining cues were visual stimuli (lines of different lengthpresented on a computer screen), and the positive andnegative responses were to touch or not touch the screen,respectively. Studies 2 and 3 used spatial location ratherthan visual or auditory cues. Rats were trained that onelocation predicted a positive stimulus (food) and anotherlocation predicted a relatively negative stimulus (no food(2) or unpalatable (quinine-flavoured) food (3)), and learntto run faster (‘go’) to the positive location and slower or notat all (‘no-go’) to the negative location. They were thentested on their speed of movement to three intermediateunreinforced probe locations with the prediction thatanimals in a more positive state would run faster to at leastsome of these. Study 5 adapted the same spatial locationtask as study (2) for dogs. Study 8 also used a go/no-go taskin which starlings were trained to flip light grey lids of foodbowls which predicted a pleasant food reward, but torefrain from flipping dark grey lids which predictedunpalatable food. Responses to probe lids of intermediateshades of grey were then observed.

Study 4 used tones as the cues, as in Harding et al.(2004), but an active choice operant response was used(e.g. if ‘positive tone’, press left lever; if ‘negative tone’,press right lever). The use of the active choice operant wasfacilitated by using food rewards for both the ‘positive’ (2pellets) and ‘less positive’ (1 pellet) reinforcers. Rats couldbe easily trained to show lever pressing for bothreinforcers, and a pilot study showed that they clearlypreferred 2 pellets to 1 pellet (Parker, 2008). Incorrectchoices following either of the training tones resulted in nofood pellets being delivered. Another innovation was to useboth ‘conventional’ single intermediate probe tones and adifferent type of ambiguous cue involving simultaneouspresentation of the two training tones. Results from bothcue types showed good agreement (Parker, 2008). Activechoice tasks were also used in study 6 in which people had

Ta

ble

2

Re

spo

nse

typ

es

an

dre

spo

nse

–o

utc

om

eco

nti

ng

en

cie

sin

jud

ge

me

nt

bia

sta

sks.

Sp

eci

es

an

dst

ud

ya

Re

spo

nse

typ

eO

utc

om

eo

fre

spo

nse

top

osi

tiv

etr

ain

ing

cue

Ou

tco

me

of

resp

on

seto

ne

ga

tiv

etr

ain

ing

cue

Ou

tco

me

of

an

y

resp

on

seto

pro

be

Po

siti

ve

Ne

ga

tiv

eP

osi

tiv

ere

spo

nse

Ne

ga

tiv

ere

spo

nse

Po

siti

ve

resp

on

seN

eg

ati

ve

resp

on

se

Ra

t(1

)Le

ve

rp

ress

No

lev

er

pre

ssFo

od

No

foo

dN

ois

eN

on

ois

eU

nre

info

rce

d

Ra

t(2

),d

og

(5)

Go

tofo

od

bo

wl

Do

n’t

go

tob

ow

lFo

od

No

foo

dN

ofo

od

No

foo

dU

nre

info

rce

d

Ra

t(3

)G

oto

foo

db

ow

lD

on

’tg

oto

bo

wl

Foo

dN

ofo

od

Un

pa

lata

ble

foo

dN

ou

np

ala

tab

lefo

od

Un

rein

forc

ed

Ra

t(4

)P

ress

left

(or

rig

ht)

lev

er

Pre

ssri

gh

t(o

rle

ft)

lev

er

2fo

od

pe

lle

tsN

ofo

od

No

foo

d1

foo

dp

ell

et

Un

rein

forc

ed

Hu

ma

n(6

)P

ress

left

(or

rig

ht)

ke

yP

ress

rig

ht

(or

left

)k

ey

Nic

eim

ag

ea

nd

ga

inp

oin

ts

No

nic

eim

ag

e

or

po

ints

ga

ine

d

Na

sty

ima

ge

an

d

lose

po

ints

No

na

sty

ima

ge

or

po

ints

lost

Ra

nd

om

ly

rein

forc

ed

Rh

esu

sm

aca

qu

e(7

)T

ou

chsc

ree

nD

on

’tto

uch

scre

en

Foo

dN

ofo

od

No

ise

No

no

ise

Un

rein

forc

ed

Sta

rlin

g(8

)Fl

ipfo

od

po

tli

dD

on

’tfl

ipli

dFo

od

No

foo

dU

np

ala

tab

lefo

od

No

un

pa

lata

ble

foo

dU

nre

info

rce

d

Sta

rlin

g(9

)P

eck

red

(or

gre

en

)k

ey

Pe

ckg

ree

n(o

rre

d)

ke

y‘i

nst

an

t’fo

od

(aft

er

1s)

No

foo

dN

ofo

od

De

lay

ed

foo

d(a

fte

r1

5s)

Un

rein

forc

ed

aS

tud

yid

en

tifi

cati

on

as

for

Ta

ble

1.

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 167

to press one of two keys on a computer keyboard depen-ding on whether they perceived a cross on a computerscreen to be closer to the ‘positive end’ of a previouslydisplayed line (if they were correct they would see apositive picture and receive a points reward, if not theywould lose points and see a negative picture. Highestpoints total won a prize), or to the ‘negative end’ of the line(a correct response would prevent the loss of points andpresentation of the negative picture, an incorrect responsewould forfeit the positive picture and points gained). Study9 also used an active choice task in which starlings weretrained to peck one key if they perceived the cue (durationof illumination of a light) to predict a positive event(immediately presented food), and another key if theyperceived the cue to predict a relatively negative event(delayed food presentation). Incorrect key pecks to thetraining cues resulted in no food delivery.

Studies used a varying number of ambiguous probecues (one (study 6); three (studies 1, 2, 3, 5, 7, 8); seven(study 9); nine intermediate and two to either side of eachtraining cue (study 4)), and in nearly all cases these wereunreinforced, apart from in study 6 where they wererandomly reinforced.

3.2.1. Generality and robustness of findings

The studies summarised in Table 1 were carried out inthree laboratories (Bristol, Newcastle, Roehampton) withfour mammalian species and one bird species, using cues ofthree different types (auditory, visual, spatial), go/no-goand active choice responses, and a variety of presumedaffect manipulations. The range of experimental contextsin which the phenomenon has been studied has bothadvantages and disadvantages. The generality and externalvalidity (i.e. applicability to other situations, populations,species, etc., Lehner, 1996) of the hypothesis that judge-ment bias (as tested in this paradigm) reflects affectivestate is strengthened if findings across a diversity ofspecies, affect manipulations, etc. support the hypothesis.This appears to be the case from Table 1 where moststudies show judgement biases as predicted. It alsoindicates that the paradigm can be used to assess affectivestates induced by both short and longer-term manipula-tions. It thus seems fair to say that judgement bias showsconsiderable promise as a new indicator of affective statein animals. It is certainly not an artefact resulting fromover-standardization of testing procedures (Wurbel,2000). However, the relatively wide-ranging combinationof species, cue modalities and affect manipulations(perhaps an inevitably of early studies in a new area)means that where results do not agree, it is difficult topinpoint the likely reasons for this, and hence to advanceknowledge systematically.

For example, how can we explain why study 4 failed tosupport the general hypothesis that animals in a putativelynegative affective state judge ambiguous stimuli nega-tively? One possibility is that the hypothesis is incorrect.However, the weight of current evidence suggests other-wise. Table 1 indicates that the clearest way in which thisstudy differed from most others was in the relativedifference between the positive and ‘negative’ reinforcersused in the judgement bias task. In this study, both

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181168

provided a food reward (i.e. the negative reinforcer was, infact, merely less positive). This was the case in order tofacilitate learning of active responses for both training cues(see earlier). However, it is possible that when there is anapparently small difference between the pay-offs of thetwo reinforcers in the task (e.g. 1 pellet vs 2 pellets), theperceived affective difference between them is small, andit thus becomes more difficult to reveal an influence ofaffect on judgement. Put another way, it is difficult to askan animal whether it predicts a better or worse event whenthe events in question differ little in their affective value(see Loewenstein and Lerner, 2003). While this may be thecase, the rats clearly preferred 2 pellets over 1 pellet(Parker, 2008), and study 9 also used food for bothreinforcers, although one reinforcer was considerablydelayed relative to the other and there is substantialtheoretical and empirical evidence to suggest that animalsshould strongly prefer more immediate rewards (seeKacelnik and Bateson, 1996). Another potential explana-tion for the failure of study 4 to reveal predicted biases isthat the unpredictable housing treatment did not induce anegative state relative to controls, and may even have hadthe opposite effect. Bodyweight data indicate that theanimals adapted to the relatively lengthy imposition (incomparison to Harding et al., 2004) of this treatment and,perhaps as a result, had in fact become less anxious andmore emboldened (as evidenced by open field and elevatedplus maze tests) and correspondingly more ‘optimistic’ inthe judgement bias task (discussed in detail by Parker,2008).

3.2.2. Differences in feeding motivation, general activity and

learning as explanations for judgement bias findings

The possibility that treatment-induced differences ingeneral activity were responsible for the findings ofHarding et al. (2004) was discussed earlier. However,any such effects would not explain the predicted biasesfound in studies where both responses required the same

active choice behaviour (e.g. studies 6 and 9). Furthermore,such effects would be likely to influence responses to thosetraining cues that require active behaviour in go/no-gotasks (e.g. run), as well as to ambiguous probe cues. Instudies 2, 3 and 5, animals in different putative affectivestates showed no differences in response to training cuesdespite showing predicted biases in their responses toambiguous probe cues, although there was a suggestionthat the active lid flipping response to the positive trainedcue was slightly diminished in unenriched birds in study 8.

The preceding points also apply to any treatment-induced effects on feeding motivation or food rewardvaluation (e.g. anhedonia) in tasks involving food reinfor-cers, except that such effects could also explain the resultsof active choice tasks. However, in study 9 where this is apossibility, although unenriched birds were more likely toperform an active key-peck predicting delayed food inresponse to the training cue associated with instant food, itseems unlikely that this was simply the result ofdiminished feeding motivation, because both active key-pecking responses were for the same sized food reinforcer(Matheson et al., 2008). Furthermore, the affect manipula-tion treatments used in study 4 were very similar to those

used in studies 1 and 2. If these treatments reliably causedfeeding motivation differences in rats, they should haveresulted in differences in responses to the training cues,especially those predicting the larger food reward, but thiswas not observed (Parker, 2008).

Overall, it appears unlikely that treatment-induceddifferences in general activity or feeding motivation canprovide a general explanation for the predicted biasesobserved in the studies in Table 1. Independent measuresof these possible effects, in particular those that can bedirectly related to the relevant operant behaviour, wouldhelp to rule them out unequivocally. Studies using activechoice tasks can usually rule out any confounding effects ofactivity differences between treatments. Studies that donot use food reinforcers would avoid any food motivationeffects, but so far such studies have been restricted tohumans (study 6) and one ongoing study of dogs at arescue shelter in Bristol.

Another possible confounding explanation for pre-dicted biases observed in go/no-go tasks where the trainedresponse to a negative cue is to ‘do nothing’, is that animalsin a negative state learn faster that the ambiguous probecues provide no reinforcement (positive or negative) andthus are quicker to stop responding (this explanation doesnot apply to the active choice tasks where a null response isnot possible). However, potentially stressful affect induc-tion manipulations or measures (e.g. studies 1, 3, 5, 7) aremore likely to result in diminished learning abilities due,for example, to distracted attention and acute or chroniceffects on memory formation and retrieval, although mild‘stress’ may enhance memory consolidation (see Mendl,1999). Similarly, environmental enrichment, which wasused to induce a relatively positive state in two of the go/no-go studies (2 and 8), is generally thought to have abeneficial effect on learning and memory abilities (e.g. vanPraag et al., 2000; Young, 2003). If both these generalfindings apply to the go/no-go studies in Table 1, we wouldpredict that, if anything, animals in a more positive statewould learn faster about the non-reinforced nature ofambiguous probe cues and hence stop responding to themsooner, rather than vice versa. Furthermore, in those go/no-go studies in which subjects were trained on thejudgement bias task while housed under affect manipula-tion conditions, and where data are available on the speedof discrimination task learning (studies 2 and 5), there wasno indication that animals in a putative negative statelearnt faster (e.g. Burman et al., 2008a).

A related possibility is that animals learn general rulesabout stimulus–outcome contingencies from the treat-ment environment in which they are housed and thislatent learning influences their responses in the judgementbias tests. For example, animals in unpredictable housingmight learn that unpredictable events are usually negative,while those in enriched housing might learn the opposite.When presented with an ambiguous (perhaps perceived asunpredictable) cue in the judgement bias test, they mightthen make the appropriate response due to a latentlearning effect. Although this is perhaps possible, it is not astrong explanation for the effects of (one-off) treatmentsthat are unlikely to involve latent learning of general rules(e.g. bright/dim light; music mood manipulation; veter-

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 169

inary inspection). Furthermore, in study 2 (Burman et al.,2008a), rats in both treatment groups experienced enrich-ment until just prior to testing (when one group hadenrichment objects removed) and yet showed clearlydifferent biases. In the studies of starlings, subjectsexperienced both enriched and unenriched environments,and in study 8 (Bateson and Matheson, 2007), birds thathad experienced enriched prior to unenriched environ-ments showed a stronger contrast between their judge-ments of ambiguity under the two conditions than thosethat received the treatments in the opposite order. If latentlearning in enriched environments that unexpected/novelevents are usually positive was an explanation of judge-ment biases in this study, one would expect the oppositeorder effect: latent learning would spill over from theenriched to the subsequent unenriched treatment (but notvice versa) rendering judgement biases more rather thanless similar across treatments, in comparison to birdsexperiencing the opposite treatment order.

3.2.3. Did manipulations really induce changes in

affective state?

Following the above arguments, judgement biasesfound in the studies in Table 1 are not easily accountedfor by treatment-induced differences in food motivation,activity or learning. However, what is the evidence that theaffect manipulations used did indeed cause the desiredchanges in affective state resulting in the observed biases?In most studies, the rationale was that the same, or similar,environmental manipulations to those used had beenshown in previous studies (including those by the authorsthemselves, e.g. Burman et al., 2006) to cause changes in anumber of commonly used behavioural and physiologicalindicators of emotional state and animal welfare. Whilethis is a reasonable argument, it would be strengthened ifindependent measures of affective state (i.e. ‘conventional’tests of emotion) confirmed these effects.

Independent measures were used alongside judgementbias tests in some studies (e.g. 1, 2, 4, 6) and producedvarying results. Responses to the PANAS affective statequestionnaire (Watson et al., 1988) confirmed that themusic manipulation used in study 6 induced the predicteddifferences in the human participants’ mood state. In study4, there was some evidence for an agreement betweenindependent measures of affect and cognitive biasperformance in that rats exposed to unpredictable housingappeared to be both less anxious, as indicated by open fieldand elevated plus maze tests, and more optimistic in thejudgement bias task (Parker, 2008). However, the judge-ment biases observed in studies 1 and 2 were notassociated with, respectively, changes in sucrose con-sumption and frequency of 50 kHz calls, both of which arethought to indicate differences in affective state (decreasedsucrose consumption may indicate anhedonia: Willner,1997; 50 kHz calls may indicate positive affect: Knutsonet al., 2002; Burman et al., 2007). There are at least twopossible reasons for this. First, the independent measuresof affective state may not have been appropriate, orsensitive enough, to detect the specific affective changes.Certainly, there is much debate as to the validity of thesucrose consumption measure (Reid et al., 1997; Phillips

and Barr, 1997), and it is also possible that an affect-induced judgement bias occurred in the absence of theanhedonic state that this measure is designed to assess.Likewise, there is some debate over the extent to which the50 kHz calls unequivocally indicate positively valencedstates (Wohr et al., 2008), and rates of calling were ratherlow in study 2 (Burman et al., 2008a). Second, theindependent measures of affect may have been providingan accurate assessment of affect, and the manipulationshad not induced changes in affective state. If so, theobserved judgement biases were either type I errors, orwere being mediated by mechanisms that did not dependon changes in affective state. While this seems unlikely, wecannot rule it out completely, and it emphasises the needto learn more about the processes mediating cognitive biaseffects in animals (see later).

Although we have made some criticisms of existingmeasures of affective state (see earlier; Paul et al., 2005), itremains important in future studies to continue investi-gating whether they co-vary with judgement bias, andhence can provide evidence of construct or convergentvalidity (strengthened when several measures of apsychological construct such as emotion agree), and alsowhether there are circumstances in which one or othertype of measure is more sensitive. Another valuableapproach is to add further to the range of affectivemanipulations that are studied, following the rationalethat the more diverse and unrelated manipulations thatcan be shown to result in predicted changes in judgementbias, the less likely that they are all doing so throughconfounding non-affective routes. In this respect, it isreassuring that manipulations which differ considerably intheir physical properties and whether they are short- orlong-term (e.g. unpredictable housing, enrichment, bright/dim light, separation anxiety score, veterinary inspection),have all resulted in predicted biases (Table 2). However,these have all been manipulations of the animal’s externalenvironment and the use of pharmacological agents (e.g.anxiolytics) to more directly manipulate the animal’s‘internal (neural) state’ would substantially strengthenthis approach. For example, it will be important to showthat pharmacological manipulations lead to the predictedjudgement bias effects (e.g. Murphy et al., 2006), and thatthe effect of an environmental manipulation on judgementbias can be reversed by the use of an (a priori) appropriatepharmacological agent, or vice versa.

3.2.4. Different types of bias? Bias location and

reinforcer value

Table 1 indicates that while studies 1, 7 and 8 detectedpredicted biases most clearly at probe locations nearest tothe positive training cue (study 7 also detected a bias at thecentral probe), studies 2 and 9 observed them closest to thenegative training cue, and the remaining studies detectedthem at the central probe (studies 5 and 6 (which only hadone probe type)), or all probe locations (study 3). Thesedifferences are potentially interesting because there issome evidence to suggest that people experiencingdifferent types of same-valence affective states may showdifferent types of cognitive bias. In particular, depressionmay be more closely linked to a decreased tendency to

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181170

expect (and recall) positive events, while anxiety isassociated with an increased anticipation (and recall) ofnegative events, although both states often co-occur inindividuals (MacLeod et al., 1997; Mineka et al., 1998;Stober, 2000; MacLeod and Salaminiou, 2001; Miranda andMennin, 2007). If the location of the ambiguous probe cueat which biases are detected can be interpreted in terms ofchanged anticipation of positive or negative events, thiscould allow judgement bias tasks to discriminate betweensame-valenced states such as anxiety and depression. Suchan interpretation is plausible if we assume that theanticipated reinforcement from a probe cue is moststrongly influenced by the reinforcer associated with thetraining cue to which the probe is most similar. Thus, aprobe cue close to the positive training cue would tend totrigger anticipation of a positive reinforcer and therefore adecreased response could be interpreted as indicatingdiminished expectation of a positive event.

However, it may be that one reinforcer is more stronglyassociated with both training cues than the other. Forexample, in a hypothetical judgement bias study involvinga positive food reinforcer and a negative electric shockreinforcer, it may be that the negative reinforcer dominatessuch that both anxious and depressed subjects perceiveboth training cues as predicting different probabilities ofthis outcome (e.g. very likely (‘danger’) vs very unlikely(‘safe’)). In this scenario, a depressed subject showing abias at a probe close to the positive cue might actually beperceiving a greater likelihood of the negative eventhappening, rather than a decreased likelihood of foodarriving. Unfortunately, we cannot rule out this possibility,and so interpretations in terms of enhanced expectation ofbad things or diminished expectation of good things needto be made with care (see Burman et al., 2008a).

This discussion also emphasises the general importanceof the affective value of reinforcers. As mentioned earlier,the relative size of the affective difference betweenpositive and negative reinforcers may influence howsensitive the test is at revealing affect-induced cognitivebiases. Furthermore, the absolute affective value of thereinforcers may influence the type of bias that is observed.For example, if the negative/less positive reinforcer is mild(e.g. no food), there may be a general tendency for allsubjects to respond to intermediate probes, and those closeto the positive training cue, as if they predict the positivereinforcer because the cost of getting it wrong andreceiving the negative reinforcer is relatively small.Affect-induced differences in response may only appearat the probes closest to the negative training cue where theperceived likelihood of the negative reinforcer occurring ishighest. On the other hand, if the negative reinforcer ismore severe (e.g. noise, unpalatable food), all subjects maybe more likely to respond cautiously to intermediateprobes and those closest to the negative training cue tominimise the chance of receiving the negative reinforcer,and hence affect-induced biases will be more likely toappear at probes nearest to the positive training cue (seeBurman et al., 2008a).

A further, technical, point is that the nature of theoperant response used in a task may also influence thelocation of biases. For example, in spatial go/no-go tasks

involving running to different cue locations (e.g. studies 2,3, 5), it is possible that subjects run at maximum speed tothe positive cue location resulting in a ‘ceiling’ effect suchthat it is not possible for them to run any faster whenexposed to a positive affective manipulation. It would thusbe difficult for any positive bias in such a task to beexpressed at probes close to the positive training cue,though it could be expressed at probes closer to thenegative cue.

Finally, it may also be possible to study specific types ofbias by choosing particular reinforcer values. For example,tests using ‘positive’ (e.g. food) and ‘neutral’ (e.g. no food,but small time/energy cost) reinforcers may be best atrevealing differences in anticipation of positive events(perhaps more relevant to depression-like states (seeearlier), or even the detection of ‘happiness’, e.g. Nygrenet al., 1996). Tests using ‘negative’ (e.g. loud noise,unpalatable food) and ‘neutral’ (e.g. no food, but smalltime/energy cost) reinforcers are likely to be technicallydifficult because subjects may simply default to perform-ing the safe negative response, but could neverthelessallow a more specific focus on biases in anticipation ofnegative events (perhaps relevant to anxiety-like states).

In summary, there is potential for useful information tobe gathered from the type of bias observed in a study, andfor tasks to be designed to investigate particular sorts ofbias. However, care must be taken to consider complicat-ing technical and interpretational issues, and the truesignificance of different types of bias remains to bedetermined.

3.3. Judgement bias: other approaches

Although judgement bias tasks following the Hardinget al. (2004) design have been the focus of most empiricalwork in this area, other approaches have been taken. Brilotet al. (2009) developed a task based on natural behaviour.Starlings were tested on their approach or avoidanceresponse to food situated close to naturally aversive eye-spot stimuli that were presented in either an unambiguousor ambiguous form. It was hypothesised that birds in amore negative affective state would be more likely to delaytheir approach to the ambiguous stimuli. The hypothesiswas not supported, perhaps because the affect manipula-tion was not fully effective and/or because eye-spots exerttheir effects through conspicuousness rather than byinducing an emotional response (Brilot et al., 2009). Parkerand colleagues (Parker, 2008; see Rowe and Skelhorn,2005) are also developing a task that capitalises on naturalbehaviour – the unlearnt avoidance by newly hatchedchicks of certain ‘warning’ colours – and investigating howaffect manipulations influence pecking at these and‘neutral’ colours. These types of task have the potentialto be much quicker to implement than the judgementbias protocols discussed above, because no training isrequired.

A different approach, based on fear conditioning, hasbeen taken by Tsetsenis et al. (2007). They trained micethat a light was a perfect predictor of impending electricshock, while a tone was a partial predictor. Responses tothese conditioned stimuli were then tested. The hypothesis

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 171

was that anxious animals would show a stronger freezingresponse to the partially predictive ‘ambiguous’ tone cuerelative to less anxious animals, but would not differ intheir response to the unambiguous light cue. They studiedknockout mice lacking the serotonin 1A receptor, whichshow enhanced anxiety-related behaviour in several tests,and found that these animals did indeed exhibit a morepronounced freezing response to the ambiguous but notthe unambiguous cue in comparison to wild-type controls.Subsequent experiments suggested that neurons in thehippocampal dentate gyrus mediated this effect, perhapsby determining the strength of association made betweencues of different predictive value and associated uncondi-tioned stimuli during the learning phase of the experiment(Tsetsenis et al., 2007; see also Nader and Balleine, 2007).

A final study that may partly be explained in terms ofbiased judgement of ambiguity was carried out by Burmanet al. (2008b). Rats were trained to run down a runway for12 pellets of food. Once running at a fairly constant speed,the size of the food reward was unexpectedly decreased tojust 1 pellet for all remaining trials. Subsequently, all ratsran more slowly to this new small reward than thoseanimals that had been trained to run for 1 pelletthroughout the study – a negative contrast effect (seeFlaherty, 1996) – but those rats which had recently beenswitched from enriched to barren housing (and hencelikely to be in a more negative affective state, see Rolls,2005) showed a longer lasting negative contrast effect thananimals which remained in the enriched environment andreceived additional enrichment objects. One explanation isthat, like humans (Wenzlaff and Grozier, 1988; Hajcaket al., 2004; Chiu and Deldin, 2007; Tucker and Luu, 2007),animals in a negative affective state are more sensitive toloss or failure (see Burman et al., 2008b). In this study, suchenhanced sensitivity was reflected in a more pronounced‘disappointment-like’ contrast response to loss of the largefood reward, and one cognitive process underlying this

Fig. 3. Highly simplified schematic diagram of some of the processes and brain are

a judgement bias task. p and n refer to cues predicting positive or negative events

text (Section 3.1) and Fig. 1 for details). Current affective state may exert modula

system, amygdala and orbitofrontal cortex (OFC) amongst others. Although a se

inter-connected and parallel processing circuits. Motivational goals may also affe

of the text.

response may have been akin to a judgement bias—unenriched rats had a lower expectation that the nowuncertain reward size would return to its previous largervalue.

Overall, these other methods for assessing judgementbias show promise, and two studies provide furthersupport for the general hypotheses relating affective stateto biased judgement of ambiguity in animals. However, itis unclear whether they are tapping exactly the samedecision-making processes as the Harding et al. test and,more generally, what these processes actually are.

3.4. Attention and memory bias

Changes in attention to, and subsequent memorystorage and retrieval of, cue-related information may beamongst the processes underlying observed judgementbiases (see below). Biases in these processes have beeninvestigated directly in humans (see earlier), but so faronly rarely in animals. A study of attention bias is beingcarried out in rhesus monkeys (E. Bethell, pers. comm.),and studies of attention and memory bias are also beingconducted in our lab. Attention bias tasks in particular maybe very quick to implement.

4. Cognitive and neural processes underlyingjudgement biases

Judgements of ambiguous stimuli are the end result of adecision-making process that involves several concep-tually distinct components: sensory registration of thestimulus; evaluation of the stimulus and likely decisionoutcome; selection of a response. The animal’s affectivestate may modulate some or all of these processes (Fig. 3).In this section, we briefly discuss whether and how suchmodulation may occur. We ground our discussion intheory and findings from neurobiological and psychologi-

as that may underlie evaluation and response to an ambiguous stimulus in

, respectively, and P and N to the appropriate responses to these cues (see

tory influences through activity in the mesocorticolimbic dopamine (DA)

quence of events is depicted, the processes are likely mediated by highly

ct these processes as described in the text. Details are provided in Section 4

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181172

cal studies of decision-making. Our aim is to identifypotential mechanisms underlying judgement bias, andexperimental approaches that could in future be used toexplore these.

From a practical animal welfare perspective it isperhaps not necessary to understand the processesunderlying judgement biases. If it can be shown empiri-cally that judgement biases are closely and reliablyassociated with affective state, especially affective valence,then this on its own is a good reason to use these measuresas indicators of animal emotion and welfare. On the otherhand, a better understanding of underlying mechanismsmay lead to more refined measurement and open upopportunities for the development of completely newtests. It may also offer new insights into affective processesand disorders in general.

Psychological and neurophysiological studies suggestthat the following events occur during the processing of acue signalling an emotionally salient outcome (reward orpunishment). Incoming sensory information is attended toand processed into a neural ‘representation’ (percept) atthe level of the primary sensory cortices (Adolphs, 2002;Rolls, 2005). It is suggested that this sensory perceptusually carries no ‘value’ information—it is an ‘objective’(or ‘what’) representation of the stimulus (Rolls, 2005).Subsequent processing appears to attribute value to thepercept and to the outcomes of behavioural responses to it(expected utility). This processing may occur in brain areassuch as the amygdala and orbitofrontal cortex and involvechanges in dopaminergic (DA) activity in the mesocorti-colimbic system (Schultz et al., 1997; Cardinal et al., 2002;Rolls, 2005; Berridge, 2007). Selection of the appropriatebehavioural action consequent on this computation of theexpected utility of a decision likely involves the basalganglia (Redgrave et al., 1999; Rolls, 2005; Bogacz, 2007).These processes are summarised in Fig. 3. Although it ishelpful to think about them as occurring sequentially, it isimportant to emphasise that they are implemented bymultiple inter-connected and parallel-processing circuitsdistributed throughout the brain. In the following discus-sion, we briefly consider how these processes mightfunction in the context of the judgement bias paradigm,and how they may be influenced by affective state.

4.1. Registration of the ambiguous stimulus: attention and

perception

Although some perceptual processes appear to be‘value-free’, this is not always the case. A large amountof sensory information impinges on an organism at any onetime and attention to this information is selective,reflecting a limit to attentional capacity. Psychologicalstudies show that negative affective states and traits,particularly anxiety, can direct attentional resources tostimuli which are associated (through learning or as aresult of ‘evolutionary preparedness’) with danger orthreat (Mathews and Mackintosh, 1998; Mogg andBradley, 1998; Bishop, 2007). Attention can thus bediverted from ongoing tasks. The neural basis of sucheffects may involve modulatory signals from the amygdalaenhancing threat-processing, and overriding those from

prefrontal cortex that support ongoing ‘neutral’ task-related processing (Mathews and Mackintosh, 1998;Bishop et al., 2004; Bishop, 2007; see also DeRubeiset al., 2008). In such circumstances, information processingcan thus be influenced by stimulus value at an early stage.

In the judgement bias studies in Table 1, ambiguousprobe cues are usually presented on their own (there is noobvious competing stimulus) and so there is littleopportunity for shifts in attention to be expressed andhence to affect resulting choices. However, threat-biasedattention processes in animals in a negative affective statemight be responsible for quicker detection of probes thatare similar to the negative training cue, resulting in morerapid responses to these. Analysis of response latencies, aswell as choice, may thus reveal treatment differences andsuggest potential underlying mechanisms (cf. Trimmeret al., 2008). Moreover, in cases where two differentlyvalenced stimuli are presented simultaneously, as duringthe test sessions involving presentation of the two trainingtones in study 4, and presentation of lines with ‘positive’and ‘negative’ ends in study 6, attention biases may directattention to one of the stimuli (e.g. the one associated witha negative outcome), and thereby influence any judgementbiases observed.

Attention bias is one example of the more generalfinding that perception, far from being the passive‘stimulus-driven’ process that it was once thought to be,appears to be a much more active and selective process.There is now considerable evidence that what is perceivedis influenced by current goals, motivational state andprevious experience and expectations (e.g. Engel et al.,2001; Rolls, 2008). This influence may be mediated by ‘top-down’ feedback pathways from brain structures ‘higher up’the sensory processing pathway (e.g. prefrontal or parietalcortex) to those lower down (e.g. thalamus), perhaps byproviding some form of ‘bias signal’ to sensorimotorcircuits (Miller, 2000). For example, perception andrecognition of emotional cues can involve the modulationof activity in primary sensory cortices by structures such asthe amygdala and orbitofrontal cortex (Adolphs, 2002).

Active sensory perception may allow an animal’scurrent motivational or emotional state to influenceexactly how it perceives an ambiguous stimulus. Forexample, dehydrated people are more likely to perceivetransparency, a property of water, in ambiguous stimulithan are non-thirsty people (Changizi and Hall, 2001).There is evidence that trait anxious people perceiveambiguous (morphed) emotional expressions as beingmore negative than non-anxious people, and that a state ofanxiety coupled with a concurrent threatening context hasthe same effect (e.g. Blanchette et al., 2007). It is thusconceivable that, for example, animals in a more negativestate might perceive and incorrectly classify a probe cue asactually being the negative training cue, hence explainingsome of the findings from Table 1. This would be mostlikely for probes that were nearest to the negative trainingcue and hence most similar to them and, more generally, ifprobes and training cues were all very similar. Some of thestudies in Table 1 used probes which are likely to havebeen easily distinguishable from training cues (e.g. spatiallocation). However, this may have not been the case for

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 173

other studies For example, the tones used in studies 1 and 4are clearly discriminable by rats when presented simulta-neously (Syka et al., 1996; Talwar and Gerstein, 1998), butmay be less so when presented with gaps of 15 s or more,as in the tests. In general, more difficult discriminationtasks involving very similar positive and negative trainingcues would enhance the potential for perceptual mechan-isms to exert an influence.

4.2. Evaluation of ambiguous stimuli: expected value and

probability of associated outcomes

Once the ambiguous stimulus has been registered, andassuming that it is not perceived as actually being one orother of the training cues, it must be evaluated. In ourjudgement bias task, it is unlikely that ambiguous cues aretreated as being completely novel and task-irrelevant asresponses shown in most of the studies in Table 1 generatesmooth generalisation curves (e.g. Fig. 2) rather than flat orerratic curves. Thus, ambiguous stimuli are probablycompared in working memory with representations ofthe two training cues and their associated response–outcome contingencies. It is likely that the animal ‘decides’between the two most salient stimulus–response–outcomecontingencies (Fig. 1): (i) the ambiguous cue is a variant ofthe positive stimulus and hence response P should beperformed to experience the positive event (with the riskthat if this decision is incorrect, the negative event will beexperienced); (ii) the ambiguous cue is a variant of thenegative stimulus and hence response N should beperformed to avoid the negative event (with the risk thatif this decision is incorrect, the positive event will be missed)(precise contingencies vary according to study design).

A prominent model of decision-making, expectedutility (EU) theory, emphasises that when decisions aremade with uncertainty about the exact outcome (i.e. mostdecisions), they should be based on a combination of theanticipated value (or magnitude) and probability ofdifferent outcomes—‘expected utility’ (see Loewensteinet al., 2008). The goal is to maximise expected utility,which is (often implicitly) assumed to be a commoncurrency proxy (cf. McNamara and Houston, 1986) for thesurvival and reproductive consequences of choices (Rolls,2005). In the judgement bias task, the factors affectingresponses to ambiguous stimuli are thus likely to be thevalue of the two key outcomes (experiencing either thepositive or negative event) to the individual, and theanticipated probability of their occurrence, which willdepend on the perceived likelihood that the ambiguous cueis a variant of either one or other of the training cues.

There is evidence that value and probability of decisionoutcomes are estimated separately in the brain (Fiorilloet al., 2003; Knutson et al., 2005; Trepel et al., 2005;Tobler et al., 2007). If so, affective states may influenceboth parameters independently as well as, or instead of,influencing an integrated representation of expected utility,and thereby bias the final decision that is made. Indeed theinfluence of ‘hot’ emotional inputs to decision-makingalongside ‘cold’ dispassionate calculations of utility isincreasingly emphasised (Loewenstein and Lerner, 2003;Loewenstein et al., 2008).

4.2.1. Stimulus or outcome value

Contemporary neuroscience and neuroeconomicsresearch suggests that the rewarding or (less studied)punishing value of a stimulus or outcome, and theassociated motivational control of actions related toacquiring or avoiding them, may be coded in a numberof inter-connected brain areas. These include corticolimbicregions such as the orbitofrontal cortex, amygdala, anteriorinsula, and anterior cingulate, the mesocorticolimbicdopamine system which includes projections from themidbrain ventral tegmental area to the basal ganglia(ventral striatum) and many of the above areas, andopioidergic systems in parts of the basal ganglia such asthe nucleus accumbens and ventral pallidum (Schultzet al., 1997; Redgrave et al., 1999; Rolls, 2005; Ernst andPaulus, 2005; Pecina et al., 2006; Berridge, 2007; Seymouret al., 2007; Knutson and Greer, 2008; Leknes andTracey, 2008). Different researchers emphasise the roleof different areas. The numerous inter-connectionsbetween areas makes it very difficult to determine whichare critical for particular valuation processes, and whichfunction as conduits for information or simply reflectactivity in other areas (see Rolls, 2005; Berridge, 2007).Given these caveats, we now very briefly describe thehypothesised role of some areas in order to illustrategeneral principles. Details will vary according to species.For example, the frontal cortex is likely to play a moreprominent role in primates than rodents.

Neurophysiological studies in primates indicate thatsensory representations of stimuli acquire a value ‘tag’ inbrain areas such as the orbitofrontal cortex and amygdala(e.g. Rolls, 2005; Padoa-Schioppa and Assad, 2006).Neuronal responses to a taste stimulus, for example,will decrease as that stimulus is consumed, parallelingthe behavioural phenomenon of specific satiety. Sensoryand physiological signals from the periphery likelymediate these motivational effects. OFC neurons firingin response to a stimulus that predicts an emotionallysalient reinforcer will also rapidly change their firing rateaccording to whether the stimulus is associated with apositive or negative outcome (Rolls, 2005; in rats, thebasolateral amygdala appears to mediate this flexibility,Cardinal et al., 2002). In contrast, neurons firing to thesame stimulus in primary sensory cortices do not showvalue-related changes, indicating that they code value-free representations of the stimuli (Rolls, 2005).

The OFC and amygdala have connections with the brainmesocorticolimbic dopaminergic system which is alsoimplicated in the valuation of reward-related stimuli andmotivation of behaviour towards such stimuli. The precisenature of this role is debated. One possibility is thatactivity in this system reflects the valuation of a predictedreward and also any mismatch between this predictionand the actual reward received (reward prediction error).This latter signal may represent a general neural substratefor reinforcement learning (Schultz et al., 1997; Montagueand Berns, 2002). Others argue that a causal role of DAactivity in this process is illusory, and that DA activity isinstead the consequence of inputs from other areas of thebrain, such as the prefrontal cortex, and hippocampus,that are involved in predicting rewards and measuring

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181174

prediction error (Rolls, 2005; Berridge, 2007). The mainrole for the DA system may instead be to attribute‘incentive salience’ to a stimulus and motivate the animalto access it (‘wanting’), while not coding for its pleasurablerewarding sensory properties (‘liking’) which aremediated by other neurotransmitter systems, notablythe opioid system, in areas such as the nucleus accumbensand ventral pallidum (Pecina et al., 2006; Berridge, 2007).A synthesis of these two ideas has also been proposed(McClure et al., 2003).

In the judgement bias task, the valuation and incentivesalience of stimulus–response–outcome contingenciesmay be mediated by these neural systems. For example,activation of unconditioned opioidergic ‘liking’ responsesto food reward will activate dopamine mediated ‘wanting’responses (Spruijt et al., 2001; Berridge, 2007) that becomeassociated with the relevant training cue and response. Thecue itself will thus come to motivate ‘wanting’ (coded asneural activity in the DA system and/or in cortical areassuch as the OFC, Rolls, 2005; Berridge, 2007), and maygenerate a reward prediction signal in the DA system.‘Disliking’ of negative stimuli may be mediated in a similarway. There is evidence that the same brain areas, and bothopioidergic and dopaminergic systems, may mediateprocesses associated with pain and pleasure. For example,striatal dopamine neurons show an inhibition of phasicfiring in response to cues signalling aversive events asopposed to enhanced firing bursts in response to predictorsof pleasant events. This system may thus provide acommon currency for valuation of both positive andnegative outcomes (Leknes and Tracey, 2008).

How might these reward-evaluation systems be influ-enced by affective state and hence alter the relativevaluation of positive or negative outcomes in a judgementbias task? Activity of the DA system appears to be sensitiveto environmental stressors and to be associated with mooddisorders such as depression (Spruijt et al., 2001; South-wick et al., 2005; Cabib, 2006; Dunlop and Nemeroff,2007), thus providing a potential link between rewardevaluation and affective state. The relationship is complexand it appears that short-term controllable aversiveexperiences may enhance DA system activity, whilelonger-term stressors and situations that the animalcannot easily control result in a decrease in DA activity(Cabib and PuglisiAllegra, 1996; Southwick et al., 2005;Cabib, 2006). This latter effect may occur due to stress orpain-induced enhancement of tonic dopamine levels thatact to attenuate phasic dopamine release and hence reduceresponses to pleasurable cues/stimuli and perhaps under-pin the anhedonic decrease in reward valuation observedin many depressed patients (Southwick et al., 2005;Dunlop and Nemeroff, 2007; Leknes and Tracey, 2008).In the judgement bias task, this may decrease the value orincentive salience of the positive stimulus–response–outcome contingency, which in turn may decrease thechances of the animal showing response P to theambiguous stimulus. Conversely, short-term enhancementof DA activity by stressors may increase reward valuationand actually promote incentive salience of the positivecontingency in mildly stressed individuals (Spruijt et al.,2001; van der Harst and Spruijt, 2008). If so, stimulus or

outcome value may be enhanced in mildly negative states,perhaps leading to potentially adaptive ‘optimistic-like’responding in judgement bias tasks, but diminished inmore pronounced or prolonged negative states. The finalresponse in a judgement bias task will, however, dependon perceived probability of outcomes as well as theirperceived value, and it may be that such probabilities co-vary monotically with affective state such that, forexample, negative state is always associated with anincrease in the perceived probability of negative events,and hence increased caution—a potentially adaptiveresponse. The links between affective modulation of valueand probability require more research and are brieflydiscussed further in Section 4.2.3.

Modulation of amygdala and OFC activity may alsoinfluence the incentive salience of both positive andnegative events (Rolls, 2005; Murray, 2007). Mood statesmay be coded as widespread changes in baseline neuronalactivity in these areas, occurring due to carry-over effectsof individual emotional events on neural activity ormodulating influences of the DA system, and hence alterneural processes that attribute value (Rolls, 2005). Theymay also bias memory retrieval. In the context ofjudgement bias tasks, this could result in ambiguous cuestriggering memory retrieval of whichever of the twostimulus–response–outcome contingencies is congruentwith current mood, and thereby influencing the valueattached to the ambiguous cue. Such mood congruentmemory effects, observed in psychological studies (Minekaet al., 1998; Lewis et al., 2005; Ramel et al., 2007), likelyinvolve signals between emotion-coding areas and mem-ory-coding areas such as the hippocampus (Rolls, 2005).

Overall, there appear to be a number of plausibleneurophysiological mechanisms by which backgroundaffective state might influence the valuation of stimu-lus–response–outcome contingencies. If, in our judgementbias task, one contingency is valued more highly than theother, or the reward of one is valued as outweighing thecosts of the other, then responses to ambiguous stimuli thatcould predict either outcome will likely be biased in favourof the more highly valued one. Affect-induced changes inthe valuation of outcomes should also lead to changes inresponses to the training stimuli that clearly predict one orother outcome (e.g. negative affect decreasing responsive-ness to the positive contingency), thus resulting in biasedresponses to both ambiguous probe cues and training cues(see Section 3.2.2).

4.2.2. Stimulus or outcome probability

Alongside decision outcome valuation, the anticipatedprobability of an outcome is the other major variable that isthought to influence decision-making. There is evidencethat aspects of outcome probability are coded separately tovalue in the brain (e.g. Critchley et al., 2001; Volz et al.,2003; Dreher et al., 2006; Tobler et al., 2007). For example,Knutson et al. (2005) showed that activity in the mesialprefrontal cortex was associated with the probability ofreward outcomes during tasks involving monetary gains orlosses. Activity in the lateral intraparietal area has beenshown to change according to the probability that a visualsaccade response will result in reward (Platt and Glimcher,

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 175

1999; Trepel et al., 2005). There is also increasing evidencethat midbrain dopaminergic activity corresponds withdiscrepancies between predicted and actual reward out-comes that are sensitive to outcome probabilities (Schultzet al., 1997; Montague and Berns, 2002). For example, theonset of cues predicting different probabilities of rewarddelivery is associated with changes in dopaminergicactivity that correlate with reward probability (Fiorilloet al., 2003). Furthermore, stimuli associated with highuncertainty of reward delivery (e.g. probability of 0.5)appear to alter the activity of dopamine neurons and thosein the OFC (Fiorillo et al., 2003; Tobler et al., 2007).

If anticipated probabilities are coded separately tovalue, this offers the opportunity for affective state tomodulate these too. In the judgement bias task, theoccurrence of smooth generalisation curves indicates thatanticipated probabilities are influenced by the perceivedsimilarity of the ambiguous cue to the two training cues.Whether and how they may be influenced by affectivestate is, however, unclear. The DA system may exertmodulatory effects as it appears to do for value, but themechanisms are uncertain because the neural substrates ofprobability coding have been less well studied. Never-theless, it remains possible that the pessimism ofdepressed and anxious people, and perhaps animals, ispartly or wholly dependent on a fundamental change intheir subjective perception of the probabilities of good andbad things happening. It does at least appear that risk-averse individuals show increased activation of brain areas(lateral OFC) which are associated with the perception ofuncertainty (variance) of an outcome (Tobler et al., 2007),indicating a link between trait characteristics and prob-ability coding.

If affect modulates probability perception, we wouldpredict that it should influence responses to the ambiguous

probe cues in our judgement bias task by altering theprobability of each outcome that the subject assigns tothese stimuli. However, for training cues which areassociated with a fixed and unvarying outcome prob-ability, we might not expect any bias of this sort. Thisdiffers from our predictions about the influence of rewardvalue modulation and, if correct, could help to distinguishbetween affective modulation of probability and value.

4.2.3. Expected utility

Why should affect modulate the valuation and antici-pated probability of outcomes separately? One reason maybe that these parameters need not always co-vary in thereal world. For example, the value of a stimulus such as aspecific food decreases as it is consumed, but theprobability of encountering it need not change in thesame way. Outcome valuation may thus be modulatedprimarily, though not exclusively, by specific motivationalstates, while the anticipated probabilities of outcomes mayperhaps be influenced by more general states, such asmoods. Interestingly, psychological studies suggest thatenhanced desirability (valuation) of an outcome does notnecessarily increase the subjective probability (optimism)of that outcome happening, indicating a dissociationbetween the two (Krizan and Windschitl, 2007). Separatemodulation of value and probability may also allow

different affective states to exert appropriate adaptiveeffects. For example, anxiety might be expected to increasethe perceived probability of negative events and hencealter alertness to threat, while depression might decreaseboth the value and probability of positive events andthereby promote a conservation and withdrawal strategy.

Of course, it may be that it is only the integratedcombination of probability and value – expected utility –that is modulated by affective state. This makes some senseif expected utility acts as a common currency for guidingdecisions (Montague and King-Casas, 2007). Neuralcorrelates of expected utility have been observed inprobabilistic decision tasks, and these occur in areas suchas the amygdala, OFC and ventral striatum which may alsocode reward value separately (Breiter et al., 2001; Rollset al., 2008) as well as in other areas such as the anteriorcingulate cortex (Knutson et al., 2005) and the ventrome-dial prefrontal cortex (Daw et al., 2006). It may be thatthese representations are subsequently modulated byaffect via similar routes to those mentioned above. Indeed,activation of the anterior cingulate cortex appears to occurwhen people are imagining positive future events and,more generally, in trait optimistic people (Sharot et al.,2007). Biases in expected utility, like those for value, wouldlikely influence responses to both ambiguous cues andtraining stimuli in judgement bias tasks.

4.3. Loss, risk and ambiguity aversion

Judgement biases may thus occur due to the influenceof affect on the attribution of value, probability andexpected utility to stimuli and related outcomes. Out-comes with particular anticipated values or probabilitiesmay also be subject to further decision-making biases.People show a general tendency to avoid losses and risky orambiguous outcomes, mainly studied in the context ofgambling tasks (e.g. Bechara et al., 1997; Shiv et al., 2005),and this has been attributed to emotional influences ondecisions (Loewenstein et al., 2008). This has led toreformulation of EU (expected utility) theory into deriva-tions such as prospect theory and ‘risk-as-feelings’ theory(Kahneman and Tversky, 1979; Loewenstein et al., 2001;Platt and Huettel, 2008; Rangel et al., 2008) that try toincorporate these emotional effects on decision-making.

Loss aversion is exemplified by the observation thatpeople will usually reject gambles that offer a 50/50 chanceof winning or losing money, unless at least twice as muchmoney can be won as lost (e.g. Tom et al., 2007;Loewenstein et al., 2008; see Kahneman and Tversky,1979). This widespread bias thus weights decisions, atleast within the dimensions that have been studied (e.g.financial reward/loss), in favour of avoiding actions thatmay result in loss. The neural correlates for such biases areunclear. Some studies report activity changes in emotion-related structures such as the amygdala when losses areanticipated or experienced (e.g. Kahn et al., 2002).Amygdala activation in more fearful/anxious people(DeRubeis et al., 2008; Krishnan and Nestler, 2008) mightthus underpin an increased loss aversion. Other studiessuggest that the relative value of gains and losses are codedasymmetrically in areas such as the ventromedial pre-

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181176

frontal cortex that receive inputs from the midbraindopaminergic system (Tom et al., 2007). Affect may thusexert an influence on this value coding through routesdiscussed earlier, and there is evidence that individualdifferences in neural sensitivity to losses are associatedwith behavioural variation in loss aversion (Tom et al.,2007). The noradrenergic system, involved in anxiety andstress responses, may also enhance loss aversion offeringanother route whereby affect might influence thisresponse bias (Rogers et al., 2004; see Trepel et al., 2005).

The enhancement of loss aversion by affective state couldcause subjects in judgement bias tasks to make decisionsthat minimise loss. In tasks where the outcomes are positive(e.g. food) vs neutral (e.g. no food), the subject does not risklosing anything that it already has, but it does risk losing apotentially available reward (or prospect). In this context,loss aversion may, counter-intuitively, increase the ten-dency to interpret ambiguous cues as positive in order tominimise chances of losing the prospect. In studies wherepositive and negative outcomes are used, predictions aremore difficult to make as the influence of affect on valuation(‘dread’) of the negative outcome will likely have a muchstronger effect on responses to the ambiguous stimuli thanthe tendency to avoid losing a prospective reward. In bothcases, however, loss aversion is unlikely to have directeffects on responses to unambiguous training stimuli.

In the decision-making literature, the concept of risk isusually applied to situations where subjects have knowl-edge of outcome probabilities and variance, and can makedecisions on this basis. Greater variance is equated withhigher risk. Ambiguity or uncertainty, on the other hand,applies to situations where they have no knowledge of thelikelihood of different outcomes. It appears that humansshow a general aversion to both risky and ambiguousoutcomes (Loewenstein et al., 2008). Behavioural ecologyresearch also indicates that animals generally tend to berisk averse with respect to increasing variation in rewardsize, but risk prone when variability is in the delay toreward. Possible mechanisms underlying these findingsare discussed by Kacelnik and Bateson (1996). Evidencethat risk aversion varies with state in animals is equivocal.Theory predicting that, as an animal’s resources dwindle tothe minimum level required for survival, risk-takingshould increase, receive mixed support (Kacelnik andBateson, 1996). In humans, increased risk aversion hasbeen linked to trait anxiety (Maner et al., 2007), but alsomore generally, and contrastingly, to positive mood (Kligerand Levy, 2003). Context may also have an importantinfluence in gambling tasks, with risk aversion beingobserved when choosing between gains, and risk-seekingoccurring when choosing between losses (Kahneman andTversky, 1979; Rogers et al., 2003). The latter effect mayreflect loss aversion. It is thus difficult to make straightfor-ward predictions as to how risk aversion may be influencedby affect and alter behaviour in judgement bias tasks.However, where there are systematic differences in thevariability of different responses in these tasks (e.g. oneresponse predicting either a positive or negative outcomewhile the other response predicts a neutral outcome; seeTable 2), it is possible that affect-related differences in riskaversion may underpin observed judgement biases.

Aversion to ambiguous (unknown) relative to risky(known) probabilities of choice outcomes has been reliablyobserved in human gambling tasks (Ellsberg, 1961;Camerer and Weber, 1992; Loewenstein et al., 2008).The neural basis of this aversion was investigated by Hsuet al. (2005) who found that activation in amygdala andOFC areas was significantly greater in ambiguous thanrisky gambles, indicating that these emotion-processingcentres may be activated by perception of ambiguity andmay mediate avoidance of such choices. It is thus possiblethat background affective state may influence this process.In the judgement bias task, ‘ambiguous’ stimuli are trulyambiguous in terms of their reinforcement history whenfirst encountered. However, if they are always paired withonly one outcome, and the animal is able to learn thisassociation during relatively infrequent probe presenta-tions, they will become less ambiguous. When an activechoice is required to an ambiguous cue in a judgement biastask, it is difficult to see how ambiguity aversion will affectthe choice made, though it may slow down the speed ofresponding. More pronounced slowing may be evident forthose cues that are perceived to be most ambiguous, forexample ‘central’ probes that are perceptually closest tothe mid-point between the two unambiguous trainingcues. When a go/no-go decision is required, it is possiblethat ambiguity aversion increases the chances of a no-go(avoid an active choice) response and thus might underlieobserved affect-modulated judgement biases. Ambiguityaversion should not affect responses to unambiguoustraining cues.

A caveat to this consideration of loss, risk andambiguity aversion is that the majority of human studiesto date have investigated reward gains and losses (oftenfinancial), while other outcomes, in particular overtlynegative ones like pain, that may be more relevant toanimals, are relatively unexplored.

4.4. Action selection

The final stage of decision-making involves actionselection. Many theorists propose that decisions for whichaction to show next are mediated by some form ofcommon currency (e.g. McNamara and Houston, 1986;Cabanac, 1992; Spruijt et al., 2001; Montague and Berns,2002; Montague and King-Casas, 2007), emphasising thatthere must be a process whereby different stimulus oroutcome evaluations can be integrated to yield one finalbehavioural decision. The role of the basal ganglia ininitiating action consequent on different types of incominginformation about relative decisions is currently animportant area of study (e.g. Redgrave et al., 1999;McHaffie et al., 2005; Daw and Doya, 2006; Bogacz,2007). In the context of perceptual decisions such asdeciding which direction a visual stimulus is moving,theoretical models suggests that ‘evidence’ for differentoutcomes is provided in the form of sensory neural inputand this evidence competes until an acceptance criterion isachieved that results in a decision to act in one way oranother (Bogacz, 2007). Some researchers believe that thisprocess takes place in cortical structures (Shadlen andNewsome, 2001). Others argue that it occurs in sub-

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 177

cortical basal ganglia structures (Redgrave et al., 1999;Bogacz and Gurney, 2007) which have the neuralarchitecture required to implement action selection inthis way (Redgrave et al., 1999).

Given that such integrating and evaluating systemsexist, it is possible that affective information can alter thebalance of ‘evidence’ required from competing inputs tofavour the outcome of one type of decision over another.This could occur through valuation processes discussedearlier resulting in the enhancement of specific inputsignals that, for example, favour actions that are mostlikely to lead to reward, or those that are most likely toprevent punishment. Interestingly, a recent study hasshown that environmental conditions that likely influenceaffective state can alter the functional organisation of oneimportant part of the basal ganglia, the nucleus accum-bens. The rostral part of the NAc in rats appears to beinvolved in motivating appetitive behaviour towardsrewarding stimuli, while the caudal part mediates fearfulavoidance behaviour. However, when rats are housed inpeaceful dark environments, fear-behaviour generatingsites recede and sites generating appetitive behaviourexpand into the caudal area. Conversely, when housed instressful, noisy environments, the opposite happens andsome rostral sites now trigger fearful behaviour (Reynoldsand Berridge, 2008). Although the precise mechanismsunderlying this re-tuning of the NAc are unclear, the studyindicates that affect-altering environments may indeedchange action selection mechanisms in the basal ganglia.

5. Conclusions

The emerging evidence that we have reviewed hereindicates that cognitive bias (specifically judgement bias)tasks show promise as new measures of animal emotionand welfare. Strengths include: a priori hypotheses thatframe the interpretation of specific outcome measures andallow tasks to be generalised across species; measurementof emotional valence rather than arousal; the potential toassess ‘optimism’ and positive emotions; and potentiallyclose correlation with reported subjective emotion (inhumans).

Judgement bias tasks comprise several components(cues, responses, reinforcers, affect manipulation; Tables 1and 2), and a number of concluding points can be madeabout these. Cues used should be counterbalanced acrosspositive and negative reinforcer contingencies, especially ifthey are not perceived in a linear fashion (e.g. tonefrequencies). Tasks using training cues that are percep-tually close together may be more likely to tap perceptualrather than evaluative (interpretative) biases (Section 4.1).Therefore, varying how perceptually similar the trainingcues are may allow one to address questions aboutunderlying mechanism. Similarly, the use of ambiguousprobe tests involving simultaneous presentation of bothtraining cues may engage attention bias mechanisms,while the more commonly used single intermediate probestimuli may be more likely to reveal evaluative biases(Section 4.1). In the majority of studies so far, three probecues have been used, allowing detection of differenceswhich are not always located at the central cue (and hence

would be missed if only one probe was used), althoughinterpretation of the location of differences is notstraightforward (Section 3.2.4). The use of unreinforcedprobe cues appears to have been successful in moststudies.

Go/no-go type responses have the advantage of beingrelatively easy to train but are more vulnerable toconfounding effects of treatment on general activity andlearning speed than are active choice responses (Section3.2.2). Active choice responses also offer clearer data on thechoice made (a ‘no-go’ response can only be classed as suchwithin a specified time window), but may be difficult totrain as responses to negative reinforcers.

The type of reinforcers used may influence thesensitivity of the test and the questions it addresses. Testsusing reinforcers that are similar in affective terms (e.g.small vs large quantity of food) may fail to reveal cognitivebiases (Section 3.2.1), while tests using clearly positive andnegative reinforcers are likely to be most sensitive as theyengage a number of affect-related influences on decision-making. However, if one or other reinforcer is particularlysalient, this may influence the shape of the resultinggeneralisation curves (Section 3.2.4). Tests using positiveand ‘neutral’ reinforcers (e.g. food vs no food) focus onchanges in anticipation of positive events that oftencharacterise depression in people, while the use ofnegative and neutral reinforcers targets changes inanticipation of negative events that are more typical ofanxiety states (Section 3.2.4). Food reinforcers have beenused in all animal studies to date but are vulnerable to theeffects of treatment on feeding motivation (Section 3.2.2).It would thus be informative to study other types ofpositive reinforcement, and this would provide furtherdata on the robustness and generality of findings.

The fact that different types of affect manipulation indifferent species appear to have yielded results in thepredicted direction (Table 1) is encouraging and argues forthe generality of the findings (Section 3.2.1). However,each type of affect manipulation is likely to be prone tospecific confounding effects and interpretations and thesemust be guarded for where possible, for example by usingindependent tests of these effects (Section 3.2.3). Thecomplementary use of existing measures of affective statecan also help assess the effects of affect manipulations, andan important area for further research is the use of non-environmental (e.g. pharmacological) manipulations thatwill provide a different type of treatment, further assessingthe robustness and generality of the phenomenon.

It is clear that a range of psychological and neurobio-logical processes are involved in decision-making of thesort tested in the judgement bias task, and hence are likelyto underlie any biases observed (Section 4). There are someapproaches that may allow one to infer the nature of theseprocesses. For example, Bayesian modelling of binarydecision-making (especially decision latency) as a cumu-lative process involving acquisition of ‘evidence’ for eachdecision option may allow one to discriminate perceptualfrom evaluative biases (Voss et al., 2008), and cortical fromfaster sub-cortically mediated decisions that can bemodelled as signal-detection processes (Trimmer et al.,2008). Biases in anticipated outcome value or probability

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181178

might be discriminated by examining whether responsesto training cues during tests are biased in the same way asresponses to ambiguous cues (perhaps indicating a bias invaluation; Sections 4.2.1 and 4.2.2). Probability biasesmight also be isolated for study by varying the relativelikelihood, rather than value, of outcomes signalled bytraining cues. More generally, probabilistic tasks could beused to specifically study ‘optimistic’ or ‘pessimistic’deviations from expected response probabilities (e.g. Yangand Shadlen, 2007; Cisek, 2007). Since responses toambiguous stimuli in the judgement bias task can beconceptualised as depending on the generalisation ofresponses to the two trained stimuli, studies of the shape ofgeneralisation curves around the training stimuli may alsobe illuminating (see Ghirlander and Enquist, 2003). Thegeneralisation response curves can be thought of asreflecting underlying sampling distributions stored inmemory (cf. Kacelnik and Bateson, 1996). Thus, experienceof an association between a cue and an outcome builds up amemory trace such that cues which are perceptually close,but not identical, to the trained cue also come to elicit thatassociation and the accompanying trained response.Increases in the perceived value and/or probability of anoutcome may result in (i) the peak of the curve at thetrained cue being heightened (an increase in expression ofthe trained response to this cue, perhaps indicating achange in valuation; Section 4.2.1), and/or (ii) the curvebroadening (the trained response is elicited to a broaderrange of cues). The decision response profile across cuesmay thus be used to indicate how sampling distributionshave been influenced by affect.

In conclusion, it is our view that cognitive bias offersmuch promise as a new indicator of animal emotion.However, research is at an early stage and, as this review hasshown, many issues remain to be addressed. Futurechallenges include further evaluation of the robustnessand generality of the phenomenon; possible refinement ofthe general valence-based hypothesis to one that incorpo-rates specific emotions or moods; development of morefocused tests to elucidate underlying mechanisms; devel-opment and validation of more rapid and practicable tests(e.g. attention bias tasks) that may be useable in the field;and investigations of trait differences in cognitive bias andtheir effects on the ability of animals to cope with challengesto their welfare.

Acknowledgements

We thank the UK Biotechnology and Biological SciencesResearch Council (BBSRC) Animal Welfare Programme, andthe Universities Federation for Animal Welfare (UFAW) forsupporting our research, and Emma Harding for her hardwork on the first project in this area. We also thank MelissaBateson, Emily Bethell, Rafal Bogacz, Rachel Casey, InnesCuthill, Amanda Holmes, Alasdair Houston, Linda Keeling,James Marshall, John McNamara, Stuart Semple, and PeteTrimmer for many stimulating discussions about affectivestates and cognitive bias. Thanks to Linda Keeling forhosting MM, ESP and OHPB for the period during whichmost of this paper was written. Thanks also to HannoWurbel for patiently waiting for it to arrive!

References

Adolphs, R., 2002. Neural systems for recognizing emotion. Curr. Opin.Neurobiol. 12, 169–177.

Australian Government, 2008. Australian Animal Welfare Strategy,revised edition. Commonwealth of Australia.

Baars, B.J., 2001. There are no known differences in brain mechanisms ofconsciousness between humans and other mammals. Anim. Welf. 10,S31–S40.

Baldock, N.M., Sibly, R.M., Penning, P.D., 1988. Behaviour and seasonalvariation in heart rate of domestic sheep, Ovis aries. Anim. Behav. 36,35–43.

Bateson, M., Matheson, S.M., 2007. Performance on a categorisation tasksuggests that removal of environmental enrichment induces ‘pessi-mism’ in captive European starlings (Sturnus vulgaris). Anim. Welf. 16,33–36.

Bechara, A., Damasio, A.R., Damasio, H., Anderson, S.W., 1994. Insensi-tivity to future consequences following damage to human prefrontalcortex. Cognition 50, 7–15.

Bechara, A., Damasio, H., Tranel, D., Damasio, A.R., 1997. Deciding advan-tageously before knowing the advantageous strategy. Science 275,1293–1295.

Bermond, B., 2001. A neuropsychological and evolutionary approachto animal consciousness and animal suffering. Anim. Welf. 10,S47–S62.

Berridge, K.C., 2003. Pleasures of the brain. Brain Cogn. 52, 106–128.Berridge, K.C., 2007. The debate over dopamine’s role in reward: the

case for incentive salience. Psychopharmacology (Berl.) 191, 391–431.

Bethell, E.J., Semple, S., Holmes, M., MacLarnon, A., 2007. The effect ofemotion state on responses to social stimuli by rhesus macaques.Primate Eye 92, 5–6.

Bishop, S., Duncan, J., Brett, M., Lawrence, A.D., 2004. Prefrontal corticalfunction and anxiety: controlling attention to threat-related stimuli.Nat. Neurosci. 7, 184–188.

Bishop, S.J., 2007. Neurocognitive mechanisms of anxiety: an integrativeaccount. Trends Cogn. Sci. 11, 307–316.

Blanchette, I., Richards, A., Cross, A., 2007. Anxiety and the interpretationof ambiguous facial expressions: the influence of contextual cues. Q. J.Exp. Psychol. 60, 1101–1115.

Bogacz, R., 2007. Optimal decision-making theories: linking neurobiologywith behaviour. Trends Cogn. Sci. 11, 118–125.

Bogacz, R., Gurney, K., 2007. The basal ganglia and cortex implementoptimal decision making between alternative actions. Neural Com-put. 19, 442–477.

Boissy, A., Manteuffel, G., Jensen, M.B., Moe, R.O., Spruijt, B., Keeling, L.J.,Winckler, C., Forkman, B., Dimitrov, I., Langbein, J., Bakken, M.,Viessier, I., Aubert, A., 2007. Assessment of positive emotions inanimals to improve their welfare. Physiol. Behav. 92, 375–397.

Bower, G.H., 1981. Mood and memory. Am. Psychol. 36, 129–148.Bradley, B.P., Mogg, K., Williams, R., 1995. Implicit and explicit memory

for emotion congruent information in depression and anxiety. Behav.Res. Ther. 33, 755–770.

Breiter, H.C., Aharon, I., Kahneman, D., Dale, A., Shizgal, P., 2001. Func-tional imaging of neural responses to expectancy and experience ofmonetary gains and losses. Neuron 30, 619–639.

Brilot, B.O., Normandale, C.L., Parkin, A., Bateson, M., 2009. Can we usestartlings’ aversion to eyespots as the basis for a novel ‘cognitive bias’task? Appl. Anim. Behav. Sci. 118, 182–190.

Burke, M., Mathews, A.M., 1992. Autobiographical memory and clinicalanxiety. Cogn. Emotion 6, 23–35.

Burman, O.H.P., AbouIsmail, U., Nicol, C., Day, M., Owen, D., Bailey, M.,Mendl, M., 2006. A multidisciplinary study of the long-term effectsof environmental enrichment on laboratory rat welfare. In: Mendl,M., et al. (Eds.), Proceedings of the 40th International Congressof the ISAE. International Society for Applied Ethology, Bristol,p. 72.

Burman, O.H.P., Ilyat, A., Jones, G., Mendl, M., 2007. Ultrasonic vocaliza-tions as indicators of welfare for laboratory rats (Rattus norvegicus).Appl. Anim. Behav. Sci. 104, 116–129.

Burman, O.H.P., Parker, R., Paul, E.S., Mendl, M., 2008a. A spatial judge-ment task to determine background emotional state in laboratoryrats, Rattus norvegicus. Anim. Behav. 76, 801–809.

Burman, O.H.P., Parker, R.M.A., Paul, E.S., Mendl, M., 2008b. Sensitivity toreward loss as an indicator of animal emotion and welfare. Biol. Lett.4, 330–333.

Burman, O.H.P., Parker, R.M.A., Paul, E.S., Mendl, M., submitted. Anxietyinduced cognitive bias in non-human animals. Physiol. Behav.

Cabanac, M., 1992. Pleasure: the common currency. J. Theor. Biol. 155,173–200.

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 179

Cabib, S., 2006. The neurobiology of stereotypy II. The role of stress. In:Mason, G., Rushen, J. (Eds.), Stereotypic Animal Behaviour. CABI,Wallingford, pp. 227–255.

Cabib, S., PuglisiAllegra, S., 1996. Stress, depression and the mesolimbicdopamine system. Psychopharmacology (Berl.) 128, 331–342.

Call, J., Carpenter, M., 2001. Do apes and children know what they haveseen? Anim. Cogn. 4, 201–220.

Camerer, C., Weber, M., 1992. Recent developments in modeling prefer-ences—uncertainty and ambiguity. J. Risk Uncertainty 5, 325–370.

Camille, N., Coricelli, G., Sallet, J., Pradat-Diehle, P., Duhamel, J.-R., Sirigu,A., 2004. The involvement of the orbitofrontal cortex in the experi-ence of regret. Science 304, 1167–1170.

Cardinal, R.N., Parkinson, J.A., Hall, J., Everitt, B.J., 2002. Emotion andmotivation: the role of the amygdala, ventral striatum, and prefrontalcortex. Neurosci. Biobehav. Rev. 26, 321–352.

Carruthers, P., 1989. Brute experience. J. Philos. 89, 258–269.Casey, R., Brooks, J., Basse, C., Burman, O., Paul, E.S., Mendl, M., 2008. The

use of ‘cognitive bias’ as an indicator of affective state in the domesticdog. In: Proceedings of the UFAW Animal Welfare Conference,Birmingham, p. 4.

Changizi, M.A., Hall, W.G., 2001. Thirst modulates a perception. Percep-tion 30, 1489–1497.

Chiu, P.H., Deldin, P.J., 2007. Neural evidence for enhanced error detectionin major depressive disorder. Am. J. Pyschiatry 164, 608–616.

Cisek, P., 2007. Neurobiology—the currency of guessing. Nature 447,1061–1062.

Clark, R.E., Squire, L.R., 1998. Classical conditioning and brain systems: therole of awareness. Science 280, 77–81.

Clore, G.L., Ortony, A., 2000. Cognition in emotion: always, sometimes ornever? In: Lane, R.D., Nadel, L. (Eds.), Cognitive Neuroscience ofEmotion. Oxford University Press, Oxford, pp. 24–61.

Coricelli, G., Dolan, R.J., Sirigu, A., 2007. Brain, emotion and decisionmaking: the paradigmatic example of regret. Trends Cogn. Sci. 11,258–265.

Critchley, H.D., Mathias, C.J., Dolan, R.J., 2001. Neural activity in thehuman brain relating to uncertainty and arousal during anticipation.Neuron 29, 537–545.

Cryan, J.F., Mombereau, C., 2004. In search of a depressed mouse: utility ofmodels for studying depression-related behavior in genetically mod-ified mice. Mol. Psychiat. 9, 326–357.

Damasio, A., 2000. The Feeling of What Happens. Vintage, London.Daw, N.D., Doya, K., 2006. The computational neurobiology of learning

and reward. Curr. Opin. Neurobiol. 16, 199–204.Daw, N.D., O’Doherty, J.P., Dayan, P., Seymour, B., Dolan, R.J., 2006. Cortical

substrates for exploratory decisions in humans. Nature 441, 876–879.Dawkins, M.S., 1990. From an animal’s point of view: motivation, fitness

and animal welfare. Behav. Brain Sci. 13, 1–61.DeRubeis, R.J., Siegle, G.J., Hollon, S.D., 2008. Cognitive therapy versus

medication for depression: treatment outcomes and neural mechan-isms. Nat. Rev. Neurosci. 9, 788–796.

Denny, E.B., Hunt, R.R., 1992. Affective valence and memory in depres-sion: dissociation of recall and fragment completion. J. Abnorm.Psychol. 101, 575–580.

Desire, L., Boissy, A., Veissier, I., 2002. Emotions in farm animals: a newapproach to animal welfare in applied ethology. Behav. Process. 60,165–180.

Desire, L., Veissier, I., Despres, G., Boissy, A., 2004. On the way to assessemotions in animals: do lambs (Ovis aries) evaluate an event throughits suddenness, novelty, or unpredictability? J. Comp. Psychol. 118,363–374.

Desire, L., Veissier, I., Despres, G., Delval, E., Toporenko, G., Boissy, A., 2006.Appraisal process in sheep (Ovis aries): interactive effect of sudden-ness and unfamiliarity on cardiac and behavioral responses. J. Comp.Psychol. 120, 280–287.

Dreher, J.C., Kohn, P., Berman, K.F., 2006. Neural coding of distinctstatistical properties of reward information in humans. Cereb. Cortex16, 561–573.

Dunlop, B.W., Nemeroff, C.B., 2007. The role of dopamine in the patho-physiology of depression. Arch. Gen. Psychiatry 64, 327–337.

Ekman, P., 1984. Expression and the nature of emotion. In: Scherer, K.R.,Ekman, P. (Eds.), Approaches to Emotion. Erlbaum, Hillsdale, NJ, pp.329–343.

Ellsberg, D., 1961. Risk, ambiguity, and the savage axioms. Q. J. Econ. 75,643–669.

Engel, A.K., Fries, P., Singer, W., 2001. Dynamic predictions: oscillationsand synchrony in top-down processing. Nat. Rev. Neurosci. 2, 704–716.

Ernst, M., Paulus, M.P., 2005. Neurobiology of decision making: a selectivereview from a neurocognitive and clinical perspective. Biol. Psychia-try 58, 597–604.

European Union, 1997. Treaty of Amsterdam. Office for Office Publicationsof the European Communities, Luxembourg.

Eysenck, M.W., Mogg, K., May, J., Richards, A., Mathews, A., 1991. Bias ininterpretation of ambiguous sentences related to threat and anxiety. J.Abnorm. Psychol. 100, 144–150.

Fernandes, C., File, S.E., 1996. The influence of open arm ledges and mazeexperience in the elevated plus-maze. Pharmacol. Biochem. Behav.54, 31–40.

Fiorillo, C.D., Tobler, P.N., Schultz, W., 2003. Discrete coding of rewardprobability and uncertainty by dopamine neurons. Science 299,1898–1902.

Flaherty, C.F., 1996. Incentive Relativity. Cambridge University Press,Cambridge.

Frijda, N.H., 1988. The laws of emotion. Am. Psychol. 43, 349–358.Garner, J.P., Meehan, C.L., Mench, J.A., 2003. Stereotypies in caged parrots,

schizophrenia and autism: evidence for a common mechanism.Behav. Brain Res. 145, 125–134.

Ghirlander, S., Enquist, M., 2003. A century of generalization. Anim. Behav.66, 15–36.

Grandjean, D., Scherer, K.R., 2008. Unpacking the cognitive architecture ofemotion processes. Emotion 8, 341–351.

Griffin, D.R., 1992. Animal Minds. University of Chicago Press, Chicago.Griffin, D.R., 1998. From cognition to consciousness. Anim. Cogn. 1, 3–16.Hajcak, G., McDonald, N., Simons, R.F., 2004. Error-related psychophy-

siology and negative affect. Brain Cogn. 56, 189–197.Harding, E.J., Paul, E.S., Mendl, M., 2004. Animal behavior—cognitive bias

and affective state. Nature 427, 312.Haselton, M.G., Nettle, D., 2006. The paranoid optimist: an integra-

tive evolutionary model of cognitive biases. Pers. Soc. Psychol. Rev.10, 47–66.

Herry, C., Bach, D.R., Esposito, F., Di Salle, F., Perrig, W.J., Scheffler, K., Luthi,A., Seifritz, E., 2007. Processing of temporal unpredictability in humanand animal amygdala. J. Neurosci. 27, 5958–5966.

Hsu, M., Bhatt, M., Adolphs, R., Tranel, D., Camerer, C.F., 2005. Neuralsystems responding to degrees of uncertainty in human decision-making. Science 310, 1680–1683.

Janczak, A.M., Andersen, I.L., Boe, K.E., Faerevik, G., Bakken, M., 2002.Factor analysis of behaviour in the porcine and murine elevated plus-maze models of anxiety. Appl. Anim. Behav. Sci. 77, 155–166.

Kacelnik, A., Bateson, M., 1996. Risky theories—the effects of variance onforaging decisions. Am. Zool. 36, 402–434.

Kahn, I., Yeshurun, Y., Rotshtein, P., Fried, I., Ben-Bashat, D., Hendler, T.,2002. The role of the amygdala in signaling prospective outcome ofchoice. Neuron 33, 983–994.

Kahneman, D., Tversky, A., 1979. Prospect theory—analysis of decisionunder risk. Econometrica 47, 263–291.

Kennedy, J.S., 1992. The New Anthropomorphism. Cambridge UniversityPress, Cambridge.

Kliger, D., Levy, O., 2003. Mood-induced variation in risk preferences. J.Econ. Behav. Org. 52, 573–584.

Knutson, B., Burgdorf, J., Panksepp, J., 2002. Ultrasonic vocalizations asindices of affective states in rats. Psychol. Bull. 128, 961–977.

Knutson, B., Taylor, J., Kaufman, M., Peterson, R., Glover, G., 2005. Dis-tributed neural representation of expected value. J. Neurosci. 19,4806–4812.

Knutson, B., Greer, S.M., 2008. Anticipatory affect: neural correlates andconsequences for choice. Phil. Trans. R. Soc. B 363, 3771–3786.

Krishnan, V., Nestler, E.J., 2008. The molecular neurobiology of depres-sion. Nature 455, 894–902.

Krizan, Z., Windschitl, P.D., 2007. The influence of outcome desirability onoptimism. Psychol. Bull. 133, 95–121.

Lang, P.J., Davis, M., Ohman, A., 2000. Fear and anxiety: animal modelsand human cognitive psychophysiology. J. Affect. Disorders 61,137–159.

Lawrence, A.B., 2008. Applied animal behaviour science: past, present andfuture prospects. Appl. Anim. Behav. Sci. 115, 1–24.

Lazarus, R.S., 1991. Cognition and motivation in emotion. Am. Psychol. 46,352–367.

LeDoux, J., 1996. The Emotional Brain. The Mysterious Underpinnings ofEmotional Life. Simon and Schuster, New York.

Lehner, P.N., 1996. Handbook of Ethological Methods, 2nd edition.Cambridge University Press, Cambridge.

Leknes, S., Tracey, I., 2008. A common neurobiology for pain and pleasure.Nat. Rev. Neurosci. 9, 314–320.

Lerner, J.S., Keltner, D., 2000. Beyond valence: toward a model of emotion-specific influences on judgement and choice. Cogn. Emotion 14,473–493.

Lewis, P.A., Critchley, H.D., Smith, A.P., Dolan, R.J., 2005. Brain mechan-isms for mood congruent memory facilitation. Neuroimage 25,1214–1223.

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181180

Loewenstein, G., Lerner, J.S., 2003. The role of affect in decision making. In:Davidson, R.J., Scherer, K.R., Goldsmith, H.H. (Eds.), Handbook ofAffective Sciences. Oxford University, Oxford, pp. 619–642.

Loewenstein, G., Rick, S., Cohen, J.D., 2008. Neuroeconomics. Annu. Rev.Psychol. 59, 647–672.

Loewenstein, G.F., Weber, E.U., Hsee, C.K., Welch, N., 2001. Risk as feelings.Psychol. Bull. 127, 267–286.

MacLeod, A.K., Salaminiou, E., 2001. Reduced positive future-thinking indepression: cognitive and affective factors. Cogn. Emotion 15, 99–107.

MacLeod, A.K., Tata, P., Kentish, J., Jacobsen, H., 1997. Retrospective andprospective cognitions in anxiety and depression. Cogn. Emotion 11,467–479.

MacLeod, C., Byrne, A., 1996. Anxiety, depression, and the anticipation offuture positive and negative experiences. J. Abnorm. Psychol. 105,286–289.

Macphail, E.M., 1998. The Evolution of Consciousness. Oxford UniversityPress, Oxford.

Maner, J.K., Richey, J.A., Cromer, K., Mallott, M., Lejuez, C.W., Joiner, T.E.,Schmidt, N.B., 2007. Dispositional anxiety and risk-avoidant decision-making. Pers. Indiv. Diff. 42, 665–675.

Marchant, J.N., Mendl, M.T., Rudd, A.R., Broom, D.M., 1995. The effect ofagonistic social interactions on the heart rate of group-housed sows.Appl. Anim. Behav. Sci. 46, 49–56.

Matheson, S.M., Asher, L., Bateson, M., 2008. Larger, enriched cages areassociated with ‘optimistic’ response biases in captive Europeanstarlings (Sturnus vulgaris). Appl. Anim. Behav. Sci. 109, 374–383.

Mathews, A., Mackintosh, B., 1998. A cognitive model of selective proces-sing in anxiety. Cogn. Ther. Res. 22, 539–560.

Mathews, A., Macleod, C., 1994. Cognitive approaches to emotion andemotional disorders. Annu. Rev. Psychol. 45, 25–50.

Mathews, A., Mogg, K., Kentish, J., Eysenck, M., 1995. Effect of psycholo-gical treatment on cognitive bias in generalized anxiety disorder.Behav. Res. Ther. 33, 293–303.

McHaffie, J.G., Stanford, T.R., Stein, B.E., Coizet, W., Redgrave, P., 2005.Subcortical loops through the basal ganglia. Trends Neurosci. 28, 401–407.

McClure, S.M., Daw, N.D., Montague, P.R., 2003. A computational sub-strate for incentive salience. Trends Neurosci. 26, 423–428.

McNamara, J.M., Houston, A.I., 1986. The common currency for behavioraldecisions. Am. Nat. 127, 358–378.

Mendl, M., 1999. Performing under pressure: stress and cognitive func-tion. Appl. Anim. Behav. Sci. 65, 221–244.

Mendl, M., 2001. Assessing the welfare state. Nature 410, 31–32.Mendl, M., Paul, E.S., 2004. Consciousness, emotion and animal welfare:

insights from cognitive science. Anim. Welf. 13, S17–S25.Mendl, M., Kurosu, G., Cuthill, I.C., Norton, V., Woodgate, J., Margetts, A.,

White, J., Paul, E.S., 2006. Studies of emotion-cognition links inhumans as a basis for developing new measures of animal emotion.In: Mendl, M., et al. (Eds.), Proceedings of the 40th InternationalCongress of the ISAE. International Society for Applied Ethology,Bristol, p. 46.

Miller, E.K., 2000. The prefrontal cortex and cognitive control. Nat. Rev.Neurosci. 1, 59–65.

Mineka, S., Sutton, S.K., 1992. Cognitive biases and the emotional dis-orders. Psychol. Sci. 3, 65–69.

Mineka, S., Watson, D., Clark, L.A., 1998. Comorbidity of anxiety andunipolar mood disorders. Annu. Rev. Psychol. 49, 377–412.

Miranda, R., Mennin, D.S., 2007. Depression, generalized anxiety disorder,and certainty in pessimistic predictions about the future. Cogn. Ther.Res. 31, 71–82.

Mobbs, D., Petrovic, P., Marchant, J.L., Hassabis, D., Weiskopf, N., Seymour,B., Dolan, R.J., Frith, C.D., 2007. When fear is near: threat imminenceelicits prefrontal-periaqueductal gray shifts in humans. Science 317,1079–1083.

Mogg, K., Bradley, B.P., 1998. A cognitive-motivational analysis of anxiety.Behav. Res. Ther. 36, 809–848.

Mogg, K., Bradley, B.P., 2005. Attentional bias in generalized anxietydisorder versus depressive disorder. Cogn. Ther. Res. 29, 29–45.

Montague, P.R., Berns, G.S., 2002. Neural economics and the biologicalsubstrates of valuation. Neuron 36, 265–284.

Montague, P.R., King-Casas, B., 2007. Efficient statistics, common curren-cies and the problem of reward-harvesting. Trends Cogn. Sci. 11, 514–519.

Murphy, S.E., Longhitano, C., Ayres, R.E., Cowen, P.J., Harmer, C.J., 2006.Tryptophan supplementation induces a positive bias in the proces-sing of emotional material in healthy female volunteers. Psychophar-macology (Berl.) 187, 121–130.

Murray, E.A., 2007. The amygdala, reward and emotion. Trends Cogn. Sci.11, 489–497.

Nader, K., Balleine, B., 2007. Ambiguity and anxiety: when a glass half fullis empty. Nat. Neurosci. 10, 807–808.

Nielsen, B.L., 1999. On the interpretation of feeding behaviour measuresand the use of feeding rate as an indicator of social constraint. Appl.Anim. Behav. Sci. 63, 79–91.

Nygren, T.E., Isen, A.M., Taylor, P.J., Dulin, J., 1996. The influence of positiveaffect on the decision rule in risky situations. Org. Behav. Hum. Decis.Process. 66, 59–72.

Oatley, K., Jenkins, J.M., 1996. Understanding Emotions. Blackwell Pub-lishing, Oxford.

Oatley, K., Johnson-Laird, P.N., 1987. Towards a cognitive theory ofemotions. Cogn. Emotion 1, 29–50.

Ortony, A., Clore, G.L., Collins, A., 1988. The Cognitive Structure of Emo-tions. Cambridge University Press, New York.

Padoa-Schioppa, C., Assad, J.A., 2006. Neurons in the orbitofrontal cortexencode economic value. Nature 441, 223–226.

Panksepp, J., 1998. Affective Neuroscience. The Foundations of Humanand Animal Emotion. Oxford University Press, New York.

Panksepp, J., 2005. Affective consciousness: core emotional feelings inanimals and humans. Conscious. Cogn. 14, 30–80.

Papp, M., Wieronska, J., 2000. Antidepressant-like activity of amisulpridein two animal models of depression. J. Psychopharmacol. (Oxf.) 14,46–52.

Parker, R.M.A., 2008. Cognitive bias as an indicator of emotional state inanimals. Unpublished PhD Thesis. University of Bristol.

Paton, J.J., Belova, M.A., Morrison, S.E., Salzman, C.D., 2006. The primateamygdala represents the positive and negative value of visual stimuliduring learning. Nature 439, 865–870.

Patrick, C.J., Bradley, M.M., Lang, P.J., 1993. Emotion in the criminal psy-chopath—startle reflex modulation. J. Abnorm. Psychol. 102, 82–92.

Paul, E.S., Harding, E.J., Mendl, M., 2005. Measuring emotional processesin animals: the utility of a cognitive approach. Neurosci. Biobehav.Rev. 29, 469–491.

Pecina, S., Smith, K.S., Berridge, K.C., 2006. Hedonic hot spots in the brain.Neuroscientist 12, 500–511.

Phillips, A.G., Barr, A.M., 1997. Effects of chronic mild stress on motivationfor sucrose: mixed messages. Psychopharmacology (Berl.) 134, 361–362.

Platt, M.L., Glimcher, P.W., 1999. Neural correlates of decision variables inparietal cortex. Nature 400, 233–238.

Platt, M.L., Huettel, S.A., 2008. Risky business: the neuroeconomics ofdecision making under uncertainty. Nat. Neurosci. 11, 398–403.

Plutchik, R., 1980. Emotion: A Psychoevolutionary Synthesis. Harper andRow, New York.

Porsolt, R.D., 2000. Animal models of depression: utility for transgenicresearch. Rev. Neurosci. 11, 53–58.

Ramel, W., Goldin, P.R., Eyler, L.T., Brown, G.G., Gotlib, I.H., McQuaid, J.R.,2007. Amygdala reactivity and mood-congruent memory in indivi-duals at risk for depressive relapse. Biol. Psychiatry 61, 231–239.

Rangel, A., Camerer, C., Montague, P.R., 2008. A framework for studyingthe neurobiology of value-based decision making. Nat. Rev. Neurosci.9, 545–556.

Redgrave, P., Prescott, T.J., Gurney, K., 1999. The basal ganglia: a vertebratesolution to the selection problem? Neuroscience 89, 1009–1023.

Reid, I., Forbes, N., Stewart, C., Matthews, K., 1997. Chronic mild stress anddepressive disorder: a useful new model? Psychopharmacology(Berl.) 134, 365–367.

Reynolds, S.M., Berridge, K.C., 2008. Emotional environments retune thevalence of appetitive versus fearful functions in nucleus accumbens.Nat. Neurosci. 11, 423–425.

Rogers, R.D., Lancaster, M., Wakeley, J., Bhagwagar, Z., 2004. Effects ofbeta-adrenoceptor blockade on components of human decision-mak-ing. Psychopharmacology (Berl.) 172, 157–164.

Rogers, R.D., Tunbridge, E.M., Bhagwagar, Z., Drevets, W.C., Sahakian, B.J.,Carter, C.S., 2003. Tryptophan depletion alters the decision-making ofhealthy volunteers through altered processing of reward cues. Neu-ropsychopharmacology 28, 153–162.

Rolls, E.T., 2005. Emotion Explained. Oxford University Pres, Oxford.Rolls, E.T., 2008. Top-down control of visual perception: attention in

natural vision. Perception 37, 333–354.Rolls, E.T., McCabe, C., Redoute, J., 2008. Expected value, reward outcome,

and temporal difference error representations in a probabilistic deci-sion task. Cereb. Cortex 18, 652–663.

Rowe, C., Skelhorn, J., 2005. Colour biases are a question of taste. Anim.Behav. 69, 587–594.

Rushen, J., 1986. Some problems with the physiological concept of ‘stress’.Aust. Vet. J. 63, 359–361.

Rushen, J., 1991. Problems associated with the interpretation of physio-logical data in the assessment of animal-welfare. Appl. Anim. Behav.Sci. 28, 381–386.

M. Mendl et al. / Applied Animal Behaviour Science 118 (2009) 161–181 181

Russell, J.A., 2003. Core affect and the psychological construction ofemotion. Psychol. Rev. 110, 145–172.

Schachter, S., Singer, J.E., 1962. Cognitive, social, and physiological deter-minants of emotional state. Psychol. Rev. 69, 379–399.

Scherer, K.R., 1984. On the nature and function of emotion: a componentprocess approach. In: Scherer, K.R., Ekman, P. (Eds.), Approaches toEmotion. Erlbaum, Hillsdale, NJ, pp. 293–317.

Scherer, K.R., 2001. Appraisal considered as a process of multi-levelsequential checking. In: Scherer, K.R., Schorr, A., Johnstone, T.(Eds.), Appraisal Processes in Emotion: Theory, Methods, Research.Oxford University Press, Oxford, pp. 92–120.

Schultz, W., Dayan, P., Montague, P.R., 1997. A neural substrate of pre-diction and reward. Science 275, 1593–1599.

Seymour, B., Singer, T., Dolan, R., 2007. The neurobiology of punishment.Nat. Rev. Neurosci. 8, 300–311.

Shadlen, M.N., Newsome, W.T., 2001. Neural basis of a perceptual decisionin the parietal cortex (area lip) of the rhesus monkey. J. Neurophysiol.86, 1916–1936.

Shafir, S., Reich, T., Tsur, E., Erev, I., Lotem, A., 2008. Perceptual accuracyand conflicting effects of certainty on risk-taking behaviour. Nature453, 917–951.

Sharot, T., Riccardi, A.M., Raio, C.M., Phelps, E.A., 2007. Neural mechan-isms mediating optimism bias. Nature 450, 102–105.

Shettleworth, S.J., 1998. Cognition, Evolution and Behavior. Oxford Uni-versity Press, Oxford.

Shiv, B., Loewenstein, G., Bechara, A., Damasio, H., Damasio, A.R., 2005.Investment behavior and the negative side of emotion. Psychol. Sci.16, 435–439.

Smith, C.A., Lazarus, R.S., 1993. Appraisal components, core relationalthemes, and the emotions. Cogn. Emotion 7, 233–269.

Smith, J.D., Shields, W.E., Washburn, D.A., 2003. The comparative psy-chology of uncertainty monitoring and metacognition. Behav. BrainSci. 26, 317–339.

Southwick, S.M., Vythilingam, M., Charney, D.S., 2005. The psychobiologyof depression and resilience to stress: implications for prevention andtreatment. Ann. Rev. Clin. Psychol. 1, 255–291.

Spruijt, B.M., van den Bos, R., Pijlman, F.T.A., 2001. A concept of welfarebased on reward evaluating mechanisms in the brain: anticipatorybehaviour as an indicator for the state of reward systems. Appl. Anim.Behav. Sci. 72, 145–171.

Stober, J., 2000. Prospective cognitions in anxiety and depression:replication and methodological extension. Cogn. Emotion 14, 725–729.

Stone, L.A., Nielson, K.A., 2001. Intact physiological response to arousalwith impaired recognition in alexithymia. Psychother. Psychosom.70, 92–102.

Syka, J., Rybalko, N., Brozek, G., Jilek, M., 1996. Auditory frequencyand intensity discrimination in pigmented rats. Hear. Res. 100,107–113.

Talwar, S.K., Gerstein, G.L., 1998. Auditory frequency discrimination in thewhite rat. Hear. Res. 126, 135–150.

Tobler, P.N., O’Doherty, J.P., Dolan, R.J., Schultz, W., 2007. Reward valuecoding distinct from risk attitude-related uncertainty coding inhuman reward systems. J. Neurophysiol. 97, 1621–1632.

Tom, S.M., Fox, C.R., Trepel, C., Poldrack, R.A., 2007. The neural basis of lossaversion in decision-making under risk. Science 315, 515–518.

Trepel, C., Fox, C.R., Poldrack, R.A., 2005. Prospect theory on the brain?Toward a cognitive neuroscience of decision under risk. Cogn. BrainRes. 23, 34–50.

Trimmer, P.C., Houston, A.I., Marshall, J.A.R., Bogacz, R., Paul, E.S., Mendl,M.T., McNamara, J.M., 2008. Mammalian choices: combining fast-but-inaccurate and slow-but-accurate decision-making systems. Proc. R.Soc. B (Lond.) 275, 2353–2361.

Tsetsenis, T., Ma, X.H., Lo Iacono, L., Beck, S.G., Gross, C., 2007. Suppressionof conditioning to ambiguous cues by pharmacogenetic inhibition ofthe dentate gyrus. Nat. Neurosci. 10, 896–902.

Tucker, D.M., Luu, P., 2007. Neurophysiology of motivated learning:adaptive mechanisms underlying cognitive bias in depression. Cogn.Ther. Res. 31, 189–209.

van der Harst, J.E., Spruijt, B.M., 2008. Tools to measure and improveanimal welfare: reward-related behaviour. Anim. Welf. 16 (S), 67–73.

van Praag, H., Kempermann, G., Gage, F.H., 2000. Neural consequences ofenvironmental enrichment. Nat. Rev. Neurosci. 1, 191–198.

Volz, K.G., Schubotz, R.I., von Cramon, D.Y., 2003. Predicting events ofvarying probability: uncertainty investigated by fMRI. Neuroimage19, 271–280.

Voss, A., Rothermund, K., Brandtstadter, J., 2008. Interpreting ambiguousstimuli: separating perceptual and judgmental biases. J. Exp. Soc.Psychol. 44, 1048–1056.

Warda, G., Bryant, R.A., 1998. Cognitive bias in acute stress disorder.Behav. Res. Ther. 36, 1177–1183.

Watson, D., Clark, L.A., Tellegen, A., 1988. Development and validation ofbrief measures of positive and negative affect: the PANAS scales. J.Pers. Soc. Psychol. 54, 1063–1070.

Wemelsfelder, F., 1997. The scientific validity of subjective concepts inmodels of animal welfare. Appl. Anim. Behav. Sci. 53, 75–88.

Wenzlaff, R.M., Grozier, S.A., 1988. Depression and the magnification offailure. J. Abnorm. Psychol. 97, 90–93.

Whishaw, I.Q., Dringenberg, H.C., Comery, T.A., 1992. Rats (rattus-norve-gicus) modulate eating speed and vigilance to optimize food-con-sumption—effects of cover, circadian-rhythm, food-deprivation, andindividual-differences. J. Comp. Psychol. 106, 411–419.

Willner, P., 1997. Validity, reliability and utility of the chronic mild stressmodel of depression: a 10-year review and evaluation. Psychophar-macology (Berl.) 134, 319–329.

Wohr, M., Houx, B., Schwarting, R.K.W., Spruijt, B., 2008. Effects ofexperience and context on 50-kHz vocalizations in rats. Physiol.Behav. 93, 766–776.

Wright, W.F., Bower, G.H., 1992. Mood effects on subjective probabilityassessment. Org. Behav. Hum. Decis. Proc. 52, 276–291.

Wurbel, H., 2000. Behaviour and the standardization fallacy. Nat. Genet. 26,263.

Yang, T., Shadlen, M.N., 2007. Probabilistic reasoning by neurons. Nature447 1075-U1072.

Yeates, J.W., Main, D.C.J., 2008. Assessment of positive welfare: a review.Vet. J. 175, 293–300.

Young, R.J., 2003. Environmental Enrichment for Captive Animals. Black-well Publishing, Oxford.

Zurita, A., Martijena, I., Cuadra, G., Brandao, M.L., Molina, V., 2000. Earlyexposure to chronic variable stress facilitates the occurrence ofanhedonia and enhanced emotional reactions to novel stressors:reversal by naltrexone pretreatment. Behav. Brain Res. 117, 163–171.