Production version To appear in KWM Fulford, M Davies, RGT Gipps, G Graham, J Sadler, G Stanghellini and T Thornton (eds), The Oxford Handbook of Philosophy and Psychiatry (Oxford: Oxford University Press, 2013). Delusion: Cognitive approaches Bayesian inference and compartmentalisation* MARTIN DAVIES AND ANDY EGAN Delusions in individuals with schizophrenia are personal-level phenomena and no account of delusion could be complete unless it included a rich phenomenological description of individuals’ experience of their delusions (Sass and Pienkos, this volume). Cognitive approaches aim to contribute to our understanding of delusions by providing an explanatory framework that extends beyond the personal level to the subpersonal level of information-processing systems (Cratsley and Samuels, this volume). At other subpersonal levels, contributions are also offered by neurobiological, neurocomputational, and psychopharmacological approaches.1 There are questions to be asked about the relationships between these different subpersonal levels (for example, about the relationship between the level of cognitive psychology and the level of neurobiology; Gold and Stoljar 1999). There are also questions about the relationship between the personal level, where description extends to phenomenology and normativity and where there are distinctive practices of rationalising explanation, and subpersonal levels of mechanistic description and explanation. According to one extreme view of this relationship, all that is literally true at the personal-level can be recast in the terms favoured by the sciences of the mind. According to the opposite extreme, what is distinctive and important at the personal level is independent from the sciences of the mind. But intermediate options, between reduction and independence, are available (inter-level interaction without reduction; Davies 2000). It is a familiar point that there seems to be an explanatory gap between the objective sciences of the mind and the subjective character of conscious experience (Nagel 1974; Levine 1983). But it would be an overreaction to this point to maintain that the science of colour vision, for example, could contribute nothing at all to our understanding of the normal or impaired experience – the ‘what it is like’ – of seeing colours. In a similar way, we can agree that phenomenological description is essential while maintaining that cognitive psychology and cognitive neuroscience can contribute to our understanding of * Our debt to the papers by Max Coltheart, Peter Menzies, and John Sutton (2010) and Ryan McKay (2012) will be evident on almost every page of this chapter. MD also acknowledges intellectual debts to Max Coltheart extending over thirty years. AE would like to thank Adam Elga, in particular, for informing and improving his thinking about compartmentalisation and Spinozan belief-formation. We are grateful to Richard Gipps, Matthew Parrott and Nicholas Shea for comments on, and conversations about, an earlier version. 1 The distinction between personal and subpersonal levels of description was introduced by Dennett (1969). The personal level of description is the folk psychological level at which we describe people as experiencing, thinking subjects and agents. It includes descriptions of conscious mental states as such and descriptions of normative requirements of rationality, for example. Personal-level descriptions figure in explanations in which actions are rationalised in terms of mental states such as beliefs and desires. Subpersonal levels, in contrast, are suited to the mechanistic descriptions and explanations of the objective sciences of the mind, such as information-processing psychology and neurobiology. (For discussion, see Davies 2000; Shea, this volume.) personal-level phenomena, including personal-level pathologies (such as addiction or hearing voices; Shea, this volume). Cognitive approaches to understanding delusions have focused first, not on the elaborated polythematic delusional systems or worlds of some individuals with schizophrenia, but on monothematic delusions – islands of delusion in a sea of apparent normality – and particularly, on monothematic delusions of neuropsychological origin. A starting point for understanding monothematic delusions is provided by Maher’s (1974, 1988, 1992) anomalous experience hypothesis: a delusion arises as a normal response to an anomalous experience. The methodology of cognitive approaches has been that of cognitive neuropsychiatry (David 1993; Halligan and David 2001); that is, the application of the methods of cognitive neuropsychology to psychiatric disorders. Thus, it has been assumed that the anomalous experience that figures in Maher’s proposed aetiology of delusions is the product of a neuropsychological deficit (Coltheart 2007, p. 1047): The patient has a neuropsychological deficit of a kind that could plausibly be related to the content of the patient’s particular delusion – that is, a deficit that could plausibly be viewed as having prompted the initial thought that turned into a delusional belief. Maher himself shares the assumption that ‘[t]he origins of anomalous experience lie in a broad band of neuropsychological anomalies’ (1999, p. 551) but, as we see in the quotation from Coltheart, cognitive approaches allow that Maher’s anomalous conscious experience might not always be essential. In some cases, the route from neuropsychological deficit to delusional belief might be wholly hidden from consciousness so that the delusional belief is ‘the first delusion-relevant event of which the patient is aware’ (Coltheart, Menzies and Sutton 2010, p. 264). Because the neuropsychological deficit is supposed to be related to the content of the delusion in some plausible way, the type of deficit (and anomalous experience, if any) will vary from delusion to delusion. The type of deficit may also vary between individuals with the same delusion if different deficits can prompt the same ‘initial thought’. Thus, cognitive approaches need to document types of neuropsychological deficit (and perhaps also types of experience) that could plausibly give rise to each of a variety of monothematic delusions, such as: Capgras delusion – ‘This [the subject’s wife] is not my wife. My wife has been replaced by an impostor’ (Capgras and Reboul- Lachaux 1923; Edelstyn and Oyebode 1999), Cotard delusion – ‘I am dead’ (Cotard 1882; Young and Leafhead 1996), Fregoli delusion – ‘I am being followed around by people who are known to me but who are unrecognisable because they are in disguise’ (Courbon and Fail 1927; de Pauw, Szulecka and Poltock 1987; Ellis, Whitley and Luauté 1994), mirrored-self misidentification – ‘The person I see in the mirror is not really me’ (Breen et al. 2000; Breen, Caine and Coltheart 2001), somatoparaphrenia – ‘This [the subject’s left arm] is not my arm’ (Halligan, Marshall and Wade 1995; Bottini et al. 2002), and the delusion of alien control – ‘Other people can control the movements of my body’ (Frith and Done 1989; Frith 1992). Candidate neuropsychological deficits have been proposed as factors in the aetiology of each of these delusions and others. But, in each case, there are examples of individuals who have the proposed deficit but not the delusion. The conclusion that is drawn from 2 this dissociation is that there must be some additional factor or factors implicated in the aetiology of delusions. In this chapter, we shall be concerned with the two-factor cognitive neuropsychological approach to understanding delusions (for early expositions, see e.g. Davies and Coltheart 2000; Langdon and Coltheart 2000; Davies et al. 2001; for recent reviews, see e.g. Aimola Davies and Davies 2009; Coltheart 2007, 2010; Coltheart, Langdon and McKay 2011; McKay 2012). Before moving on, we shall illustrate the proposal of a neuropsychological deficit as a first factor, and the dissociation argument for a second factor, in the widely discussed case of Capgras delusion. We shall draw on important early work in cognitive neuropsychiatry (Ellis and Young 1990), which in turn built on a well-supported model of normal face processing (Bruce and Young 1986). In the Bruce and Young model, information about known faces is stored in face recognition units (FRUs), one for each known face. When a known face is seen, one FRU will be activated to a high level and biographical information stored in a corresponding personal identity node (PIN) – such as information about the person’s occupation – will be accessed, as will the person’s name. An important functional difference between FRUs and PINs is that only a seen face will activate an FRU, whereas a PIN can be accessed from the person’s seen face or heard voice, or in other ways. In some individuals with severely impaired face recognition (prosopagnosia), skin conductance responses continue to discriminate between familiar and unfamiliar faces – there is covert recognition (Tranel and Damasio 1985, 1988; see also Bauer 1984). So, although the primary face-recognition system is damaged in these individuals, there must be a preserved connection between an early stage of face processing (the FRUs) and the autonomic nervous system. Ellis and Young (1990) proposed that the neuropsychological deficit in Capgras delusion is the mirror image of the deficit in prosopagnosia with covert recognition. In Capgras delusion, the primary face-recognition system is intact but the connection between the FRUs and the autonomic nervous system is damaged. Normally, the seen face of a loved one, such as the spouse, causes activity in the autonomic nervous system and the experience of the loved one’s face has a strong affective component. But now, with the connection between the face-recognition system and the autonomic nervous system disrupted, this component of the experience is missing. Ellis and Young’s proposal about the neuropsychological deficit and anomalous experience in Capgras delusion made a clear empirical prediction that the skin conductance responses of individuals with Capgras delusion would not discriminate between familiar and unfamiliar faces. This prediction was subsequently confirmed in four studies using photographs of familiar (famous or family) faces and unfamiliar faces (Ellis et al. 1997, 2000; Hirstein and Ramachandran 1997; Brighetti et al. 2007) – an ‘exemplary vindication’ of the new discipline of cognitive neuropsychiatry (Ellis 1998). Thus, it is plausible that a neuropsychological deficit, disconnection of the primary face- recognition system from the autonomic nervous system, is a factor in the aetiology of Capgras delusion. But there is a dissociation between this deficit and the delusion. There are individuals (patients with damage to ventromedial regions of frontal cortex; Tranel, Damasio and Damasio 1995) whose skin conductance responses do not discriminate between familiar and unfamiliar faces, but who do not have Capgras delusion (or any other delusion). There is also a report of an individual who (following temporal lobe 3 surgery for relief of epilepsy) had an anomalous ‘Capgras-like’ experience of her mother – ‘she was different, something was different about her … you can look different by, you know, doing your hair or whatever, but it wasn’t different in that way … it didn’t feel like her’ (Turner and Coltheart 2010, pp. 371–2) – but did not have the Capgras delusion.2 Thus there must be a second factor in the aetiology of Capgras delusion – presumably, in cases of neuropsychological origin, a second deficit. 1. The two-factor framework for explaining delusions Coltheart (2007 p. 1044) has proposed that, in order to explain any delusion, we need to answer two questions. First, where did the delusion come from? Second, why does the patient not reject the belief? The leading idea of the two-factor framework for explaining delusions (Coltheart 2007, 2010; Coltheart et al. 2011) is that the two factors will provide answers to these two questions. The first question is always: where did the delusion come from? – that is, what is responsible for the content of the delusional belief? The second question is always: why does the patient not reject the belief? … – that is, what is responsible for the persistence of the belief? (Coltheart 2007 p. 1044) Factor 1 is what is responsible for the belief having occurred to the person in the first place … : this factor determines the content of the delusional belief. Factor 2 is responsible for the failure to reject the hypothesis despite the presence of (often overwhelming) evidence against it … this factor determines the persistence of the delusional belief. (Coltheart 2010, p. 18) The first question is, what brought the delusional idea to mind in the first place? The second question is, why is this idea accepted as true and adopted as a belief when the belief is typically bizarre and when so much evidence against its truth is available to the patient? (Coltheart et al. 2011, p. 271) The two-factor framework has provided explanations (at least in outline) of a range of delusions including those that we have already mentioned and also anosognosia for motor impairments (Davies, Aimola Davies and Coltheart 2005; Aimola Davies et al. 2009, Aimola Davies and Davies 2009). It has been proposed that the second factor is the same in all cases of delusion (or at least in all cases of delusion of neuropsychological origin) and that it consists in an impairment of normal processes of belief evaluation, associated with pathology of right lateral prefrontal cortex (e.g. Coltheart et al. 2011, p. 285). The nature of the putative task of belief evaluation suggests that the second factor could be an impairment of executive function or working memory (or both), consistent with its proposed neural basis (Aimola Davies and Davies 2009). But the cognitive nature and neural basis of the second factor have not been specified as precisely as the nature and basis of putative first factors. 2 This patient was studied by Nora Breen and Mike Salzberg. 4 1.1 Adoption and persistence: Two options for the two-factor framework It is important to notice that the three quotations listed earlier leave open two possible interpretations of the second question (that is, the question to which the second factor is supposed to provide an answer). In the third quotation, the second question is about adoption of the delusional belief. But it is possible initially to adopt a belief and then, on reflection, to reject it. Sometimes we initially believe what we see and then realise that we are subject to an illusion or we initially believe what we are told and then realise that our informant is unreliable. In the case of a delusional belief we can ask why the belief, once adopted, is not subsequently rejected. Why is it ‘firmly sustained’ (American Psychiatric Association 2000, p. 821), why does it persist? This is more like the version of the second question that is posed in the first two quotations. Explaining a delusion requires answers to both the adoption question and the persistence question. In principle, it might turn out that two factors (two pathologies or departures from normality) are needed to answer the adoption question and that a third factor is needed to answer the persistence question. That would, of course, be incompatible with the two-factor framework, but compatible with a less specific multi- factor framework. An exactly-two-factor account must say either: (A) that no pathology or departure from normality beyond the first factor is needed to answer the adoption question and the second factor answers the persistence question; or else: (B) that two factors are needed to answer the adoption question and no additional pathology or departure from normality is needed to answer the persistence question. (In principle, it might be that option (A) is correct for some delusions and option (B) for others.) According to option (A), we should expect each dissociation of the ‘first deficit without delusion’ form (e.g. ventromedial frontal damage without Capgras delusion; Tranel et al. 1995) to be a case in which the delusional belief is initially adopted, but does not persist. According to option (B), we should expect each dissociation to be a case in which the first deficit is present, but the delusional belief is not even initially adopted. This presents a potential problem for option (A) because there is no evidence that patients with ventromedial frontal damage, for example, initially adopt the Capgras delusion but subsequently reject it (nor is there evidence that this is not the case; see Coltheart et al. 2010, p. 281 and McKay 2012, pp. 341–2, for discussion). On the other hand, there are reports of individuals who, after recovering from a delusion, still feel the attraction of the belief that they now reject. For example, a patient (HS) who had recovered from anosognosia reported that the idea that he could move his paralysed limbs still seemed credible even though he was able to reject it (Chatterjee and Mennemeier 1996, p. 227): E: What was the consequence of the stroke? HS: The left hand here is dead and the left leg was pretty much. 5 HS: (later): I still feel as if when I am in a room and I have to get up and go walking . . . I just feel like I should be able to. E: You have a belief that you could actually do that? HS: I do not have a belief, just the exact opposite. I just have the feeling that sometimes I feel like I can get up and do something and I have to tell myself ‘no I can’t’. Turner and Coltheart describe a patient in the early stages of recovery from Capgras delusion (2010, p. 371): I’ve started going through it, and seeing what could possibly happen and what couldn’t happen. That was wrong, that couldn’t happen. Even though it has happened it couldn’t. Mary couldn’t suddenly disappear from the room, so there must be an explanation for it. … And then I worked it out and I’ve wondered if it’s Mary all the time. It’s nobody else. In summary, the standard examples of deficit without delusion, which figure in the dissociation argument for a second factor, are potentially problematic for option (A) and fit option (B) better. But the examples of recovery from delusion fit option (A) well, on the assumption that the recovery resulted from remission of the second factor. Thus, not only the cognitive nature and neural basis of the second factor, but also – and even more importantly – its role in the aetiology of delusions, requires further specification.3 1.2 Bayesian approaches One of the aims of cognitive neuropsychology is to understand disorders of cognition in terms of theories or models of normal cognition. Cognitive impairments are understood in terms of damage to one or more components of the normal cognitive system. When the methods of cognitive neuropsychology are applied to delusions – pathologies of belief – what is required is an information-processing model of the normal formation, evaluation, and revision of beliefs. Thus, one of the problems faced by cognitive neuropsychiatry – in comparison with the cognitive neuropsychology of face recognition, for example – is that we do not have an articulated, still less a computationally implemented, model of normal believing. Indeed, there may be reasons of principle why it is difficult to understand believing in terms of the computational theory of mind (Fodor 1983, 2000; see Cratsley and Samuels, this volume, on Fodorian pessimism). More than twenty-five years ago, Hemsley and Garety suggested a strategy for making progress in the absence of a model of normal believing (1986, p. 52): ‘A normative theory of how people should evaluate evidence relevant to their beliefs can 3 In this chapter, we shall be defending a version of option (A), but we do not offer a resolution of the potential problem associated with option (A). In the specific case of the ventromedial frontal patients, it might be suggested that they do not initially adopt the Cagras delusion because they do not, in fact, have exactly the same neuropsychological deficit as Capgras patients (see Ellis and Lewis 2001). That suggestion provides a response to the potential problem for option (A) but at the price of removing the standard dissociation argument for a second factor. 6 provide a conceptual framework for a consideration of how they do in fact evaluate it.’ Their specific proposal was to begin from a probabilistic analysis of hypothesis evaluation and then to investigate whether individuals with delusions deviate from the normative Bayesian model. In pursuing this strategy and interpreting its results, it is important to distinguish the normative from the normal; it is important not to forget that, as Hemsley and Garety put it, there is ‘“normal” deviation from the prescriptive model’ (p. 55). Recently, the Bayesian approach has been married with the neuropsychological deficit approach in continuing development of the two-factor framework for explaining monothematic delusions (Coltheart et al. 2010; McKay 2012). A second body of work has adopted a Bayesian approach – and, specifically, the theoretical framework of predictive coding and prediction error signals, in which neural processing aims to minimise prediction error or ‘free energy’ (Friston 2005, 2009, 2010; Friston and Stephan 2007) – to delusions in schizophrenia (Corlett, Frith and Fletcher 2009; Fletcher and Frith 2009).4 In this chapter, we shall focus on the Bayesian two-factor approach to explaining monothematic delusions and on the idea that delusions arise through a process of Bayesian inference or updating. 1.3 Bayesian inference On a Bayesian approach, probabilities are updated on the basis of evidence, E, so that the new or posterior probability of a hypothesis, H, is equal to the old or prior conditional probability of H given E. This updating procedure is known as simple conditionalisation. By Bayes’ theorem, the conditional probability, P(H|E), can be further unpacked to give: Simple conditionalisation P´(H) = P(H|E) = P(H). P(E|H)/P(E). (Here, P´ is the new distribution of probabilities.) The notions of prior and posterior probabilities are relative. The prior probability of H is prior only to the evidence E; it already takes account of antecedently available evidence – today’s priors are yesterday’s posteriors. In simple conditionalisation, the evidence is treated as certain: P´(E) = 1. A more general updating procedure, Jeffrey conditionalisation (Jeffrey 1983), allows that the evidence may be less than certain, so that P´(E) < 1.5 The posterior probability of a hypothesis, H, updated on the basis of evidence E by simple conditionalisation, is proportional to the prior probability of H, P(H), and to the probability of E given H, P(E|H), also known as the likelihood of H on E. The likelihood provides a measure of how well H predicts E. In this chapter, we shall usually be more interested in the balance of probabilities between two competing hypotheses than in the 4 See Frith 2007, chapters 4 and 5, for an accessible introduction to the predictive coding approach and Corlett et al. (2007, 2009) for prediction error and delusion. See also Shea, this volume, for discussion of prediction error signals. 5 In Jeffrey conditionalisation, the probability of H is updated to: P´(H) = P(H|E). P´(E) + P(H|not-E). P´(not-E). 7 precise probability of each hypothesis. If we are considering two hypotheses, H and H , 1 2 then the ratio of posterior probabilities (the posterior odds) is the product of two other ratios, the ratio of prior probabilities (the prior odds) and the likelihood ratio: Bayes Ratio Formula P´(H ) = P(H |E) = P(H ) . P(E|H ) 1 1 1 1 P´(H ) P(H |E) P(H ) P(E|H ) 2 2 2 2 Thus, on a Bayesian approach, the balance of probabilities between two candidate hypotheses, updated on the basis of evidence E, depends on (a) how probable each hypothesis is in the light of available evidence other than E – given by the prior probability P(H) – and (b) how well each hypothesis predicts the evidence – given by the i likelihood P(E|H). i In the next two sections, we shall review two versions of the Bayesian two-factor approach (Coltheart et al. 2010; McKay 2012) in some detail. Section 2 is about the initial adoption of a delusional belief and section 3 is about the persistence of the belief. Before moving on, however, we note that there are complex and difficult issues surrounding the relationship between, on the one hand, Bayesian inference or updating and, on the other hand, abductive inference or inference to the best explanation (Lipton 2004). Coltheart et al. (2010) sketch two models of abductive inference, the logical empiricist model based on an understanding of explanation as logical implication and the Bayesian model based on a probabilistic account of explanation (p. 271): ‘the hypothesis H explains observations O to the degree x just in case the probability of O given H is x’. They adopt a Bayesian model of abduction, but we are not committed to the view that Bayesian inference is a model of inference to the best explanation. One reason is that the likelihood, P(E|H), is not in general a good measure of how well a hypothesis H explains evidence E. A hypothesis about barometer readings (e.g. the barometer is falling) does not explain weather patterns (e.g. a storm is coming), however high the likelihood (that is, the probability of the weather patterns given the barometer readings) may be. Rather, causation and explanation run in the opposite direction, from weather patterns to barometer readings (van Fraassen 1980, p. 104). More generally, Lipton’s account of inference to the best explanation takes account, not only of whether a candidate explanatory hypothesis is the most probable given the available evidence, but also of whether it exhibits explanatory virtues such as parsimony, scope, depth, unifying disparate phenomena, and making new predictions. The question then arises whether it could ever be rational to accept an explanation because of its virtues, if an alternative explanation was more probable (van Fraassen 1989). Lipton (2004) aims to neutralise this concern about the relationship between inference to the best explanation and Bayesianism by suggesting that explanatory virtues are a guide to probability, but we take no stand on that issue. In our discussion of delusions and Bayesian inference, it will be the standard Bayesian apparatus of probability assignments, likelihoods, and updating that bears the theoretical load. The notion of explanatory virtue will play only a peripheral role, in that it may influence the psychological accessibility of hypotheses. It is true that some of the literature that we shall engage with is couched in terms of inference to the best 8 explanation. But this seems to be largely inessential and, because of the issues that we have just mentioned, potentially distracting. Most or all of the theoretical work in explaining the adoption and persistence of delusional beliefs in terms of Bayesian abductive inference could be done just as well by talking about Bayesian inference simpliciter, thereby sidestepping those complex and difficult issues. 2. Bayes in the two-factor framework: Adoption of the delusional belief Coltheart and colleagues (2010) propose that the answer, in outline, to the question where a delusion came from is that it arose through a process of Bayesian inference. In principle, this might be a process of inference carried out consciously by the person with the delusion, but Coltheart and colleagues focus on the case of unconscious inferential processes. To illustrate their approach, Coltheart and colleagues provide a worked example of how Bayesian inference could lead from a neuropsychological deficit to the initial onset of Capgras delusion. 2.1 From deficit to delusional belief: Capgras delusion Suppose that, as the result of a stroke, a patient suffers disconnection of the primary face processing system from the autonomic nervous system (while the two disconnected systems themselves remain intact). Before the patient suffered the stroke, the appearance of his wife caused activation of the face recognition unit for the wife (FRU ), which W normally led to activation, not only of the corresponding personal identity node (PIN ), W but also of the autonomic nervous system. As a result of the learned association between the appearance of the patient’s wife and activation of his autonomic nervous system, the appearance of his wife generated an unconscious prediction of activity in the autonomic nervous system, and this prediction was reliably fulfilled. Following the stroke, some things remain the same and some things are different. When the patient sees his wife, the face recognition unit FRU and the personal identity node PIN are still activated, and W W activity in the autonomic nervous system is still predicted. But, because of the disconnection, the prediction is not fulfilled. The abnormal absence of the predicted autonomic activity, resulting from the neuropsychological deficit (disconnection), stands in need of explanation. The aim of the Bayesian approach is to show that a delusional hypothesis may be initially adopted as a belief as a result of Bayesian inference or updating on the basis of abnormal data, D (in this case, the absence of activity in the autonomic nervous system). Consequently, the next stage of Coltheart and colleagues’ (2010) worked example involves two competing hypotheses. One is the true hypothesis, H , that the woman that W the patient sees in front of him, who looks like the patient’s wife and says that she is the patient’s wife is, indeed, his wife. The other is the delusional hypothesis, H , that the S woman is not the patient’s wife but a stranger.6 6 Coltheart et al. (2010) do not consider hypotheses that are incompatible with both H and H , such as the W S hypothesis H , that the person that the patient sees in front of him is aunt Agatha, or the hypothesis H , that A B the person that the patient sees in front of him is Bob the bank teller. As a result of this simplification, H is S treated as the negation of H . W 9 What needs to be shown is that the ratio of posterior probabilities, P(H |D) / P(H |D) S W could favour the stranger hypothesis, H , over the wife hypothesis, H . The Bayes ratio S W formula tells us that this ratio is equal to the product of the ratio of prior probabilities and the likelihood ratio. So, how might the balance between those two ratios favour H over S H ? The prior probabilities, P(H ) and P(H ), are prior only to the to-be-explained W W S abnormal data D. They take account of antecedently available evidence including, in particular, the evidence that the woman that the patient sees in front of him looks just like his wife and says that she is his wife. Consequently, the probability that the woman is the patient’s wife is much higher than the probability that she is a stranger and the ratio P(H ) / P(H ) is correspondingly low. In contrast, Coltheart and colleagues say, the S W likelihood ratio, P(D|H ) / P(D|H ) is high: ‘It would be highly improbable for the S W subject to have the low autonomic response [D] if the person really was his wife, but very probable indeed if the person were a stranger’ (2010, p. 277). According to the worked example, then, the ratio of prior probabilities favours the wife hypothesis, H , but the likelihood ratio favours the stranger hypothesis, H . W S Whether the ratio of posterior probabilities favours H or H depends on the relative W S values of these ratios and, specifically, on whether the likelihood ratio is sufficient to outweigh the ratio of prior probabilities. Suppose, for example, that the prior probabilities favoured H in the ratio 100:1 but the likelihoods favoured H in the ratio 1000:1. Then W S the posterior probabilities would favour H in the ratio 10:1. If these were the only two S hypotheses to consider, their probabilities would be P(H ) = 0.91 and P(H ) = 0.09. S W Thus, Bayesian inference might lead from the abnormal data D to the assignment of a high probability to the hypothesis that the woman who looks just like the patient’s wife and also claims to be the patient’s wife is not his wife but a stranger, and so an impostor. Equally, if the prior probabilities favoured H in the ratio 100:1 but the likelihoods W favoured H only in the ratio 10:1, then the posterior probabilities would be reversed: S P(H ) = 0.09 and P(H ) = 0.91. S W Coltheart et al. suggest that the likelihood ratio does outweigh the ratio of prior probabilities (2010, p. 278): The delusional hypothesis provides a much more convincing explanation of the highly unusual data than the nondelusional hypothesis; and this fact swamps the general implausibility of the delusional hypothesis. So if the subject with Capgras delusion unconsciously reasons in this way, he has up to this point committed no mistake of rationality on the Bayesian model.7 One difficulty in evaluating this suggestion about Bayesian inference is that it is somewhat unclear which probabilities are to figure in the worked example. Are they supposed to be, for example, realistic probabilities or the patient’s subjective probabilities 7 Coltheart and colleagues move from the fact that the delusional hypothesis, H , provides a much more S convincing explanation of the data than the wife hypothesis, H , does to the claim that the likelihood ratio W strongly favours H . It is worth noting, however, that explanatoriness is not always a good indicator of S likelihood. From the fact that a hypothesis, H, utterly fails to explain evidence, E, it does not follow that the likelihood, P(D|H), is close to zero. 10