Auditory hallucinations and verbal spatial source monitoring: evidence from healthy adults

Auditory hallucinations and verbal spatial source monitoring:

evidence from healthy adults

Zoë Firth

69 Natal Road, London, N11 2HT, United Kingdom

A Master’s thesis submitted in partial fulfilment of the requirements for the degree of Master of Science in Clinical Linguistics

at the Joint European Erasmus Mundus Master’s Programme in Clinical Linguistics (EMCL+)

University of Groningen University of Potsdam University of Eastern Finland

Supervised by Professor Margaret Niznikiewicz (Harvard Medical School) & Dr. Srđan Popov (University of Groningen)

2 Table of Contents Table of Contents ... 2 List of Tables ... 3 List of Figures ... 3 1. Introduction ... 4

The source monitoring account of auditory verbal hallucinations ... 4

Spatial source monitoring ... 6

Source monitoring: the role of emotional valence ... 7

Hallucination-proneness ... 10

The present study ... 12

2. Methods ... 13

Participants ... 13

Materials ... 14

2.2.1 Source monitoring stimuli ... 14

2.2.2 Hallucination-proneness ... 15 Procedure ... 15 Data analysis ... 17 3. Results ... 19 Hallucination-proneness ... 19 Accuracy ... 19

Discriminability and bias ... 20

3.3.1 Demographic effects ... 20 3.3.2 Item memory ... 21 3.3.3 Source memory ... 21 4. Discussion ... 24 Limitations ... 27 Future Research ... 27 Conclusions ... 29 References ... 30

Appendix I: stimulus words ... 37

Appendix II: Launay-Slade Hallucination Scale-Revised ... 39

3 List of Tables

Table 1: Demographic distribution of participants by age and gender ……..………..… 13

Table 2: Source attribution accuracy rates by condition and emotional valence ………. 19

List of Figures Figure 1: Graphic illustration of the experimental procedure ………..…… 16

Figure 2: Frequency of hallucination-proneness scores ………...… 19

Figure 3: Item discriminability plotted by age category ...………... 20

Figure 4:Source discriminability plotted by age category ……….. 20

Figure 5: Item discriminability as a function of emotional valence ………. 21

Figure 6: Source discriminability as a function of HP and emotional valence ……..……….. 22

Figure 7: Source bias as a function of HP and emotional valence ..……… 23

4 1. Introduction

Schizophrenia is a debilitating psychiatric disorder present in approximately 1% of the population (APA, 2013). One of the most common symptoms of schizophrenia is auditory hallucinations, wherein an individual perceives an auditory stimulus in absence of any external event to which that perception can be attributed. Auditory hallucinations are experienced by between 60 and 75% of people with schizophrenia (Nayani & David, 1996). The majority of auditory hallucinations are perceived as voices (McCarthy-Jones et al., 2014), and are therefore called auditory verbal hallucinations (AVHs). These AVHs are often highly distressing and can cause significant social and functional impairment (see McCarthy-Jones, 2012); understanding the psycholinguistic and cognitive mechanisms underlying AVHs is therefore of primary clinical as well as theoretical importance.

The source monitoring account of auditory verbal hallucinations

A prominent account of AVHs is the source monitoring account, wherein AVHs are attributed to a pathological deficit in monitoring the source of self-generated speech (see Brébion et al., 2000). Under the original source monitoring framework of Johnson and colleagues (1993), a source monitoring deficit can manifest along two dimensions, namely: discriminability, the ability to make correct source attributions for perceptual events; and bias, an increased likelihood of making attribution errors in one direction compared to the other. Specifically, under the source monitoring account it is posited that AVHs are the result of a source monitoring externalisation bias, such that inner, self-generated speech is misattributed to an external source. The source monitoring account of AVHs has been tested in studies utilising the verbal source monitoring paradigm, wherein participants are presented with words that they either produce themselves, as self-generated speech (SGS), or are produced by another person, as other-generated speech (OGS); thus, the agentic source (self vs. other) of language production is manipulated. Later, participants are presented with these stimulus items as well as new items, and are asked to identify whether the item is new or old, and for old items to identify the source (i.e., self or other).

In support of the source monitoring account, studies conducted under the verbal source monitoring paradigm have shown that people with schizophrenia (PwS) who experience AVHs (PwS-AVH+) generate more self-to-other agentic externalisation errors than both healthy controls (Brébion et al., 2000) and PwS who do not experience AVHs (Brunelin et al., 2006). This externalisation bias has been corroborated by studies conducted under other behavioural paradigms, such as the auditory signal detection paradigm (Bentall & Slade, 1985a). Likewise,

5

studies of online verbal source monitoring have also yielded evidence for an externalisation bias. Within this paradigm, participants say words out loud and are simultaneously played acoustic feedback of SGS or OGS speaking that word, with or without acoustic distortion. Many online verbal source monitoring studies (e.g., Johns et al., 2001) have shown that PwS-AVH+ are more likely to identify distorted feedback of their own voice as OGS than controls and PwS without the AVH symptom. Neuroimaging studies have revealed further insight into this source monitoring externalisation bias; for example, Allen and colleagues (2007) found that while for controls and PwS without the AVH symptom, OGS elicited more activation than SGS in the left superior temporal gyrus, speech source did not modulate activity in this area amongst PwS-AVH+. Given the role of the left superior temporal gyrus in language and auditory perception, this result indicates that the enhanced auditory processing associated with external speech, and hence the distinction between external and internal speech sources, is eroded amongst PwS-AVH+. Abnormal modulation of functional connectivity patterns by speech source has also been linked to AVHs, in particular connections between language-relevant fronto-temporal regions and the anterior cingulate cortex (Mechelli et al., 2007) as well as the thalamus (Kumari et al., 2010). In this way, the externalisation bias revealed by studies conducted under the verbal source monitoring paradigm, as well as neuroimaging studies of verbal source monitoring, have supported the source monitoring account of AVHs. It is important to note that the claim that AVHs can be causally attributed to a source monitoring deficit has been disputed. In particular, the specificity of source monitoring deficits to PwS-AVH+ is not consistent, as such deficits have not only been attested to in PwS without AVHs, but in some cases were found to be correlated with non-AVH symptoms such as Schneiderian delusions (Anselmetti et al., 2007). Nevertheless, some researchers have noted a greater impairment in verbal source monitoring amongst PwS who do experience AVHs than those who do not (Brunelin et al., 2007), and a meta-analysis from Brookwell and colleagues (2013) of source monitoring and auditory signal detection studies comparing patients with and without AVH yielded a moderate to large effect size, indicating an increased externalising bias amongst the PwS who do experience AVHs. It is therefore possible that there are diagnosis-general impairments to source monitoring in schizophrenia, but that AVHs manifest as a symptom-specific source monitoring deficit targeting the language system. Comparing source monitoring across modalities, as well as triangulating research conducted across a variety of methodologies, will yield greater insight into the role of verbal source monitoring as a potential causal mechanism of AVHs. The present study will reflect on the issue of source monitoring as a causal mechanism for AVHs by examining the relationship between verbal source

6

monitoring performance and the subclinical experience of auditory hallucinations (‘hallucination-proneness’; see section 1.4) amongst healthy adults.

Spatial source monitoring

In addition to agency, another cognitive dimension of AVHs which has been investigated under the source monitoring account is spatial source, that is, whether a cognitive event is perceived inside or outside the mind. As some researchers have noted (McKague et al., 2012; Stephane, 2019), the agentic and spatial source dimensions of language production are often confounded within the methodology employed to test verbal source monitoring, for example by employing paradigms wherein generating inner speech and hearing OGS are contrasted (e.g., Brunelin et al., 2007). However, studies of the phenomenological characteristics of AVHs have revealed that agentic and spatial source are independent of each other (Stephane et al., 2003; McCarthy-Jones et al., 2014). Moreover, the characteristics of an individual’s AVHs may be related to their source monitoring performance. In a further phenomenological investigation of AVH characteristics in schizophrenia, Stephane (2019) found that not only were the agency and spatial source characteristics of participants’ AVHs independent of each other, but also that these cognitive dimensions of participants’ AVHs were associated with their error pattern in a verbal source monitoring task. That is, participants who experienced AVHs as OGS were more likely to make self-other agency externalisation errors in the verbal source monitoring task, but not other-self errors, relative to those who experienced AVHs as SGS (i.e., as their own voice). Likewise, those participants who experienced AVHs in outer space were more likely to make inner-outer spatial source externalisation errors, but not outer-inner errors, relative to those who experienced AVHs in inner space. Furthermore, spatial source bias did not differentiate internal-versus-external agency AVH patient groups, and vice versa. These results indicate that spatial source should be investigated as an independent dimension of AVHs and verbal source monitoring, and that such investigations have the potential to reflect on the causal mechanisms of AVH generation.

Traditionally, the spatial source dimension of verbal source monitoring has been isolated by comparing reading silently and aloud, that is, covert versus overt reading. While using covert reading as a form of inner speech is limited inasmuch as ‘it is not really clear where a self-generated voice, be it covert or overt speech, is actually perceived’ (McKague et al., 2012:509), the manifestation of inner speech during reading, or ‘inner reading voice’, has been validated as a phenomenological experience occurring in the majority of people (Vilhauer, 2017). Moreover, it has been demonstrated that self-generated inner speech is processed in many similar ways to overt SGS, for example being accompanied by a neural corollary

7

discharge signal (Whitford et al., 2017; Jack et al., 2019). Hence, under a spatial source monitoring paradigm which employed reading, Franck and colleagues (2000) found that PwS-AVH+ falsely recognised covertly-read material as overtly-read material more often than control participants, but did not find a difference in overt-covert error rates, thus indicating the presence of a spatial source externalisation bias amongst PwS-AVH+. However, in contrast to the agentic source monitoring bias which is usually found to be unidirectional (i.e., self-other externalisation) amongst PwS-AVH+, the spatial source monitoring bias may not be unidirectional; Stephane and colleagues (2010) found that PwS-AVH+ were more likely than control participants to make both inner-outer and outer-inner attribution biases on a spatial source monitoring task which employed reading. This result indicates that in contrast to the agentic source monitoring deficit which manifests as an externalising bias, the spatial source monitoring deficit may manifest only as a discrimination deficit: that is, a difficulty discriminating between internal and external sources, as opposed to a bias in either direction.

Source monitoring: the role of emotional valence

Verbal source monitoring performance in PwS-AVH+ is also moderated by the emotional valence of the verbal material. In particular, it has been reported that negative material is more subject to an externalising bias than neutral or positive material. Costafreda and colleagues (2008) found that symptomatic PwS produced more external attribution errors for negative than neutral words, while no such effect of emotional valence was present amongst either PwS in remission or control participants. An externalisation bias for negative material has also been demonstrated in studies of online verbal source monitoring (Johns et al., 2001), as well as voice identity processing; Pinheiro and colleagues (2016) found that PwS were less accurate than controls in recognising SGS with negative content, and that recognition of undistorted negative SGS was negatively associated with a measure of hallucination severity. This result is corroborated by electrophysiological evidence that PwS displayed increased late-stage processing for SGS compared to OGS for negative words only, while control participants showed this increased SGS processing for both positive and negative words (Pinheiro et al., 2017). Thus, the source monitoring deficit underlying the generation of AVHs appears to be particularly sensitive to language with negative content, as reflected in an increased externalising bias for negative material.

Given this behavioural evidence, many authors have related the greater externalisation bias for negative material to the phenomenological characteristics of AVHs in schizophrenia. That is, AVHs often manifest as commanding or derogatory content (Nayani & David, 1996), and AVHs with negative content are more common than AVHs with positive content

8

(McCarthy-Jones et al., 2014); hence, AVH generation could represent a process whereby the source of unwanted thoughts with negative content is externalised to reduce distress (see Morrison, Haddock, & Tarrier, 1995). However, while this account provides an intuitive explanation for the negative externalisation bias, it is important to note that not all studies have found a specific effect for negative material. An externalisation bias for positive but not negative words compared to neutral words has been reported for PwS-AVH+ (Morrison & Haddock, 1997) and individuals experiencing first-episode psychosis (Bendall, Jackson, & Hulbert, 2011); in the former study, negative words were actually related to a higher rate of internalisation (i.e., other-self) errors in the PwS-AVH+ sample. It is therefore possible that the moderating effect of emotional valence targets both positive and negative emotional content.

As such, there have been other proposals for the dissociations between neutral and emotionally valent material in source monitoring paradigms. Morrison, Haddock, and Tarrier (1995) have proposed that AVHs are externalised as a mechanism of mitigating the anxiety caused by an inconsistency between a high prevalence of emotionally arousing intrusive thoughts and specific meta-cognitive beliefs (e.g., ‘Not being able to control my thoughts is a sign of weakness’). This explanation is supported by evidence that PwS-AVH+ score more highly on meta-cognitive beliefs such as negative beliefs about the uncontrollability of thoughts (Lobban et al., 2002). Meta-cognitive processes are also implicated by evidence that PwS-AVH+ provide lower ratings of both internality and control for their source attribution responses to emotional compared to neutral words (Baker & Morrison, 1998), and that this difference is greater for words applied to the self as opposed to another individual (Ensum & Morrison, 2003). It has also been suggested that the effect of emotional valence on source monitoring may be attributable to emotional deficits which disrupt the regular memory encoding processes during which a stimulus is bound to its source information. Indeed, both trait and state emotional characteristics have been linked to AVH generation (see Waters et al., 2012), and PwS exhibit a variety of behavioural deficits in emotional processing, for example delayed processing of emotionally valent material (Seok et al., 2006). However, these deficits are present across many PwS and not only those with the AVH symptom; thus, the explanatory power of this proposal may be limited. Further research is therefore needed to disentangle whether both positive and negative material are similarly subject to source monitoring deficits in schizophrenia, and whether these effects are specific to PwS-AVH+.

Understanding how emotional valence influences psycholinguistic processes such as verbal source monitoring also has the potential to reveal the neurolinguistic deficits underlying

9

AVHs in schizophrenia. That is, a variety of research has demonstrated that PwS exhibit a disruption to the typical left-hemisphere lateralisation for language processing, for example, a reduced right ear advantage during dichotic listening tasks (Hugdahl et al., 2008). Likewise, neuroimaging studies have demonstrated a correlation between the severity of AVHs and decreased lateralisation of language by both structural and functional measures (Oertel et al., 2010). Hence, a prominent theory of AVHs in psychosis is that they are attributable to the reduced lateralisation of typical language processing pathways (Mitchell & Crow, 2005). In addition to abnormal left hemisphere processing, there is also evidence for disruptions to the lateralisation of emotional stimuli to the right hemisphere in schizophrenia, for example in the processing of emotional prosody. Not only do PwS display a behavioural deficit in interpreting emotional prosody (see Hoekert et al., 2007), but PwS-AVH+ specifically exhibit an absence of a left ear advantage for emotional prosody compared to those who do not experience AVHs (Alba-Ferrara et al., 2013), a deficit which is also reflected in electrophysiological measures (Pinheiro et al., 2014). There is thus evidence that the interaction of language and emotion is reflected in hemispheric dominances, and that a disruption of this process may result in the generation of AVHs.

Under an abnormal lateralisation framework, the specificity of source attribution deficits to negative content could be accounted for by a modified version of the Valence Hypothesis (see Killgore & Yurgelun-Todd, 2007), which posits that while all emotional content recruits the right hemisphere to some extent, negative content generates a greater degree of right hemisphere activation than positive content. In support of this hypothesis, Prete and colleagues (2020) found a right ear advantage for auditory verbal imagery with a positive valence, but no hemispheric differences for imagery with a negative valence; the authors attribute this result to the recruitment of the right hemisphere for negative content ‘cancelling out’ the left hemispheric specialisation for language processing. Evidence such as this which supports the Valence Hypothesis would explain the increased externalisation bias for negative material, as an altered ability to recruit right-lateralised negative emotion processing mechanisms would disrupt the usual source-encoding of inner speech attribution.

The phenomenological characteristic of AVHs as emotionally valent thus aligns with the behavioural and neurophysiological evidence for a violation of expected lateralisation effects for the processing of emotional linguistic material, which may specifically target negative material. It may be that emotional content, including semantic valence as well as vocal features such as prosody, mediates the relationship between speech production and source identification; hence, a deficit in schizophrenia of the processing of emotionally valent

10

language increases the likelihood of a disruption in this relationship, resulting in an increased likelihood of AVHs. In this way, investigating the effect of emotional valence on verbal source monitoring contributes to our understanding of the neurolinguistic deficits underlying AVH generation.

Hallucination-proneness

Although the majority of auditory hallucinations are associated with a psychotic disorder such as schizophrenia, AVHs also occur in members of the healthy population. While figures vary widely (see Beavan et al., 2011), it is estimated that 10 to 15% of healthy adults sometimes or regularly experience AVHs (Sommer et al., 2010), with a lifetime prevalence of AVHs reported at around 8% (Kråkvik et al., 2015). Evidence that many of the same risk factors associated with AVHs in schizophrenia have also been attested to amongst healthy adults with higher proneness to experience hallucinations has generated a continuum model of psychosis, wherein it is posited that psychotic experiences, including AVHs, are not a categorically distinct phenomenon but rather occur along a spectrum (van Os et al., 2009). Under this model, ‘hallucination-proneness’ (HP) is a variable designed to capture the experience of AVHs at a subclinical level in members of the healthy population. HP can be measured by questionnaires such as the Launay-Slade Hallucination Scale (Launay & Slade, 1981).

In support of a continuum model of psychosis, the externalisation bias in agentic source monitoring paradigms attested to in PwS-AVH+ has also been found in healthy adults with higher compared to lower HP (see Brookwell et al., 2013; cf Alderson-Day et al., 2019). Moreover, as with PwS-AVH+, the externalisation bias associated with HP appears to be sensitive to the emotional content of verbal material; Larøi and colleagues (2004) found that healthy participants with higher HP committed more external misattribution errors than low-HP participants on positive and negative but not neutral words, and this externalising bias amongst high-HP participants was significantly greater for negative than positive words. This result has been interpreted as indication that, like clinical AVH presentation, HP is associated with an externalising bias which targets emotionally valent material – especially negative material – over neutral material. This interpretation accords with the finding that HP is predicted by the presence of specific meta-cognitive beliefs related to, for example, confidence in one’s own cognitive processes (Larøi & van der Linden, 2005). Hence, the source externalisation of negative verbal material in high-HP adults may be attributable to similar cognitive mechanisms as in PwS-AVH+.

However, similar to PwS-AVH+, the precise directionality of the effect of emotional valence on verbal source monitoring in healthy adults with HP has not been entirely consistent.

11

For example, Kanemoto and colleagues (2013) found no effect of emotional valence on attribution performance in the Deese-Roediger-McDermott memory paradigm amongst high-HP participants, while low-high-HP participants’ discrimination performance increased for positive words and decreased for negative words. Hence, while the affective modulation of verbal self-monitoring appears to be present in individuals with sub-clinical HP, further research is required to elucidate whether there is a general effect of emotional valence, or whether negative words are selectively targeted.

Also similar to research conducted with PwS, the spatial source dimension of verbal source monitoring has been very rarely investigated amongst the healthy population. Noting the experimental confound between agentic and spatial source within previous methodologies, McKague and colleagues (2012) used an acoustic mannequin to create auditory stimuli that, when played to participants, sounded as if they were generated internal or external to the participant’s head. The authors found no evidence for a spatial source externalising bias amongst participants with high or low HP, nor an interactive effect of HP and emotional valence on source memory accuracy. However, this experimental paradigm isolated the spatial source distinction between two other-generated speech sources. Given that AVHs are by definition generated cognitive events, employing a methodology which utilises self-generated language to isolate the spatial source variable may hold more validity. As such, Garrison and colleagues (2017) conducted a spatial source monitoring experiment with healthy adults using a covert-overt reading paradigm, and found no association between HP score and either rate or bias of misattribution errors, although all stimuli were neutral. Together, the results of these two studies indicate that spatial source may not be a salient cognitive dimension of AVH-like experiences in the healthy population, but methodological factors necessitate further research.

Studying healthy adults with varying levels of HP is informative for a number of reasons. Firstly, it provides an opportunity to isolate the ‘trait’ characteristics associated with AVHs (see Ditman & Kuperberg, 2005). That is, the prevalence of characteristics associated with clinical AVHs amongst adults with high HP who are not psychotic indicates that these are trait characteristics associated with AVH, and therefore may be implicated in the causative mechanisms of AVH generation. However, there may also be state characteristics associated with actively hallucinating which influence source monitoring in schizophrenia. Johns and colleagues (2006) compared PwS who were currently experiencing AVHs to PwS with a history of AVHs but who were not currently experiencing them and found that current hallucinators were more likely to externally misattribute distorted SGS than previous

12

hallucinators. Hence, comparing the manifestation of processes such as source monitoring between low- and high-HP individuals may reveal important insight into the trait characteristics associated with AVH, and thus the potential causative mechanisms of AVH generation.

Secondly, the phenomenological characteristics of AVHs in clinical and non-clinical populations implicate a differential role of emotional valence between these two groups. Specifically, comparisons of clinical and non-clinical AVH-hearers have confirmed that the emotional valence of AVHs is the most significant individual predictor of the presence of psychosis amongst AVH-hearers, such that non-clinical AVH-hearers report fewer negative and more positive AVHs than AVH-hearers with a psychotic disorder (Daalman et al., 2010; Daalman et al., 2011). Examining whether and how emotional valence modulates source attributions for language in high-HP individuals may reveal important insight into how speech processing mechanisms are disrupted in the generation of clinical AVHs, which in turn may be used to inform therapies which target the AVH characteristics that are clinically significant.

The present study

While previous studies of verbal spatial source monitoring amongst healthy participants did not find evidence for a relationship between the spatial source dimension and HP, these studies did not vary the emotional valence of the stimuli (Garrison et al., 2017) or tested the spatial source dimension of speech which was not self-generated (McKague et al., 2012). The present study is designed to address this gap in the literature by testing the effect of emotional valence on the spatial source dimension of verbal source monitoring in healthy adults. Healthy adults will perform a verbal spatial source monitoring task in which they produce words covertly and overtly, and their recall for the source of these words will be tested later along with new distractor items; the spatial source of self-generated verbal material is thus manipulated using a covert-overt reading contrast. Participants will also complete a questionnaire measure of HP. This study is thus designed to address the following research questions:

RQ1) Are healthy adults with greater proneness to experience auditory hallucinations more likely to make externalising errors in a verbal spatial source monitoring paradigm? Expectation: The source monitoring account predicts that adults with greater HP will exhibit a greater rate of external misattribution errors than low-HP adults. However, given the bi-directional bias found by Stephane and colleagues (2010), it is possible that high-HP adults will exhibit higher rates of both types of bias, and hence the source monitoring deficit will manifest as lower discriminability overall, with no bias in either direction.

13

RQ2) Does emotional valence have an effect on the likelihood of making a spatial source monitoring error amongst healthy adults?

Expectation: In accordance with Larøi and colleagues (2004), it is expected that positive and negative words will generate more misattribution errors (i.e., lower discriminability) than neutral words; negative words may in turn generate more misattribution errors than positive words. It is also possible that this source monitoring deficit will manifest specifically as a greater externalisation bias to emotional words. However, in accordance with the first research question, these effects, if present, are only expected in high-HP individuals.

2. Methods Participants

Participants were 46 adult, native speakers of English between the age of 18 and 65 years old (n(female) = 31, n(male) = 13, n(non-binary) = 2). Table 1 below shows the demographic distribution of participants by age category and gender. The exclusion criteria for participation were: a history of either a developmental language disorder (e.g., dyslexia) and/or a psychotic disorder (e.g., schizophrenia, schizotypal personality disorder, bipolar disorder). All participants reported completing or having completed a university-level degree at the undergraduate or graduate level.

Table 1: Demographic distribution of participants by age and gender

Age Gender 18-25 26-35 36-45 46-55 56-65 Total Female 10 11 3 4 3 31 Male 3 3 2 0 5 13 Non-binary 1 1 0 0 0 2 Total 14 15 5 4 8 46

Participants were recruited via word-of-mouth and online advertisement. Participation in the experiment was not financially reimbursed, but participants could contact the researcher to request to be entered into a voucher lottery. Participants gave informed consent before taking part in the study, and ethical approval was provided by the Research Ethics Committee of the Faculty of Arts, University of Groningen (CETO).

14 Materials

2.2.1 Source monitoring stimuli

Stimuli for the source monitoring experiment were 144 adjectives. Part of speech was verified using the SUBTLEX-UK database (van Heuven et al., 2014), and only words wherein the ‘dominant’ (i.e., most frequent) part of speech for that lemma was ‘adjective’ were used. In accordance with the norming procedure of Warriner and colleagues (2013), only those words which over 70% of participants evaluated as being known to them in Kuperman and colleagues (2012) were eligible for stimulus selection, as the validity of the ratings yielded by words which are familiar to more participants is considered higher. Potentially trauma-related words (e.g., suicidal, alcoholic) were removed from the dataset.

Stimulus words were divided into three valence categories, of 48 words each. To derive the valence categories, valence ratings were taken from the Glasgow Norms database (Scott et al., 2019); ratings in this database are made along a scale from 1 to 9. The total range for average valence ratings amongst eligible adjectives within this database was between 1.35 and 8.58; therefore, this range was divided into three approximately equal sections to yield three valence categories: below 3.8 (negative), 3.8 to 6.2 (neutral), and above 6.2 (positive). Accordingly, t-tests revealed that the negative word list had a lower average valence rating (M = 2.67) than the neutral (M = 5.06) and positive (M = 7.10) word lists (t(91.95) = -17.80, p < .001; t(92.25) = -38.21, p < .001), and the neutral word list also had a lower average valence rating than the positive word list (t(87.11) = -16.09, p < .001).

To minimise the influence of lexical variables which have been shown to influence psycholinguistic processing, the three valence lists were matched for the following variables: frequency; age-of-acquisition; imageability; concreteness; familiarity; and phonological complexity, operationalised as length in syllables and length in phonemes. Values for length in syllables and in phonemes were derived from the N-Watch database (Davis, 2005); values which were not available in N-Watch were input manually. Frequency values were taken from the SUBTLEX-UK database, and were log-transformed for the norming analysis. The normed ratings for age-of-acquisition, imageability, concreteness, and familiarity were also taken from the Glasgow Norms database.

However, the three valence lists were not matched for arousal ratings (which were also taken from the Glasgow Norms database); previous research has yielded a consistent relationship between arousal and valence, wherein positive and negative words are more arousing than emotionally neutral words (see Warriner et al., 2013). Hence, in this study, the list of neutral words had a significantly lower average arousal (M = 4.42) than both positive (M

15

= 5.18) and negative (M = 5.05) word lists (t(88.46) = -3.84, p < .001; t(91.52) = 3.09, p = .003, respectively). The positive and negative word lists did not significantly differ by arousal ratings (t(93.26) = -0.71, p = .483).

Hence, the stimuli for the source monitoring experiment were 144 adjectives, with 48 appearing in the inner speech experimental condition of the source monitoring task, 48 appearing in the outer speech condition, and 48 words employed in the recall task as unseen distractors. Within each condition list (inner, outer, recall), there were 16 words of each of emotional valence (positive, negative, neutral). The experimental and recall lists were matched for the same 7 psycholinguistic variables as the valence lists. See Appendix I for a full list of items.

2.2.2 Hallucination-proneness

Hallucination-proneness was measured using a subset of items from the revised version of the Launay-Slade Hallucination Scale (LSHS-R; Bentall & Slade, 1985b) derived by McCarthy-Jones and Fernyhough (2011); this subset of items has been employed in previous studies of HP in healthy adults (Garrison et al., 2017; Alderson-Day et al., 2019). The revised scale comprises nine items total: five related to hallucination-like experiences in the auditory modality, and four in the visual modality. Each item is rated on a 4-point Likert Scale from 1 (‘never’) to 4 (‘almost always’). See Appendix II for a full list of items. Factor analyses have revealed independent loadings of the auditory and visual items, and the revised scale has been found to have an internal reliability of .73 (McCarthy-Jones & Fernyhough, 2011); this is higher than the internal reliability of the original revision of the LSHS-R reported by Morrison, Wells, and Nothard (2000). Participants completed all nine items of this scale, but only the five auditory items are included in the present analysis. Hence, possible scores ranged from 5 to 20 points.

Procedure

The study was presented through the online experiment-hosting platform PCIbex (Zehr & Schwarz, 2018). Figure 1 displays a graphic illustration of the experimental procedure. First, participants completed the source monitoring task, which consisted of two stages: stimulus presentation and recall. Both the stimulus presentation and recall stages of the source monitoring task began with task instructions and four practice trials. To reduce potential order effects, there was a mandatory 3-minute break between the stimulus presentation and recall stages, during which participants passively viewed images of fractals. Examinations of the time stamps on the experimental portions of the task confirmed that all participants completed the stimulus presentation and recall stages within 42 minutes (M = 23.08; SD = 4.80).

16 1500 ms

OUTER

+

happy

In which condition of the experiment did you encounter this word?

happy

Task: None; passively looking at images Stimuli: 12 images of fractals

Trial duration: 15s per image; no inter-trial interval Overall duration: 3 minutes

QUESTIONNAIRE

Task: Say stimulus word out loud if pre-trial cue is OUTER, say

stimulus word ‘inside mind’ if pre-trial cue is INNER

Stimuli: 96 items (48 OUTER words, 48 INNER words) Trial duration: 4.5s per trial (see right), 1s inter-trial interval

Task: Rate the extent to which each questionnaire item is personally true, from 1 (‘never’) to 4 (‘almost always’)

Stimuli: 9 questionnaire items Trial duration: Unlimited

Task: Indicate via key-press which condition each stimulus word was presented in (INNER, OUTER, or NEW WORD i.e., unseen distractor) Stimuli: 144 items (48 INNER words, 48 OUTER words, 48 NEW words)

Trial duration: 7s per item or until valid key-press; 1s inter-trial interval Task instructions + practice trials

I hear a voice speaking my thoughts aloud.

Unlimited duration Task instructions + practice trials

Task instructions Task instructions

SEQUENCE TASK SPECIFICATION TRIAL STRUCTURE

Figure 1: Graphic illustration of the experimental procedure

15 seconds

F = INNER J = OUTER

17

During stimulus presentation, words were presented individually in sequential trials at an inter-trial interval of 1000ms. At the beginning of each trial, the trial condition was signalled by the appearance of the word ‘INNER’ or ‘OUTER’ at the centre of the screen for 1500ms, followed by a fixation cross for 1000ms. After the fixation cross, the stimulus word appeared on the screen for 2000ms. To provide an adjuvant perceptual cue to trial type, the condition text was presented in different colours (‘INNER’ in blue and ‘OUTER’ in green). Text was presented centrally in the Verdana font at size 36.

During recall, items were presented in sequential trials at an inter-trial interval of 1000ms. For each item, participants were asked to indicate via key-press which condition of the experiment they thought they had encountered the word in: INNER, OUTER, or new item. Participants had seven seconds to make a decision for each item; if a valid keypress was made within seven seconds, the next trial began automatically. No feedback was given. As in the stimulus presentation stage, item and keypress text were presented centrally in size 36 Verdana font.

For the 96 target items, two lists were constructed such that item-condition presentation was counterbalanced across all participants; all participants saw the same 48 recall items. Hence, for the stimulus presentation stage, items were split into 3 blocks of 32 items, administered in a Latin Squares design across participants. For the recall stage, items were split into 4 blocks of 36 items, administered in a Latin Squares design across participants. In both stages, item order was pseudo-randomised, such that no more than two items of the same valence or three items of the same condition were presented sequentially. In both the stimulus presentation and recall stages, participants had the option to take a short break between each block of stimuli.

Finally, participants completed the 9-item sub-scale of the LSHS-R. To incentivise completion of all sections of the study, participants were given the option to see their score on the memory task after the questionnaire.

Data analysis

Statistical analysis was conducted in R and RStudio (R Core Team, 2020; RStudio Team, 2020). Following previous analyses of HP and source monitoring (e.g., McKague et al., 2012; Garrison et al., 2017), signal detection theory was used to calculate discriminability and bias of both item memory and source memory (see Macmillan & Creelman, 2005). Discriminability is a measure of a participant’s accuracy in discriminating the source of two types of information (old-new, inner-outer), and bias is a measure of whether a participant is more likely to make a discrimination error in one direction compared to the other. Hence, d’ was calculated to

18

measure discriminability, and c was calculated to measure bias. Outcome measures were calculated as follows (see Stanislaw & Todorov, 1999):

d’ = z(hits) – z(false alarms) c = -0.5[z(hits) + z(false alarms)]

To calculate these outcome measures, the z scores associated with the rates of ‘hits’ and ‘false alarms’ are used. For item memory, hits were classified as old items correctly recognised as old (inner or outer), and false alarms were new items incorrectly recognised as old. For source memory, hits were classified as inner items correctly recognised as inner, and false alarms were outer items incorrectly recognised as inner. Source memory d’ and bias scores were calculated only from items correctly recognised as old. To avoid scores of positive or negative infinity, minimum and maximum rates (e.g., 0 false alarms) were corrected with a 0.5 increase or decrease, respectively.

Thus, a d’ score of 0 represents chance performance; for item memory, a maximum d’ score would be 4.02, while for source memory, a maximum d’ score would be 3.73. Likewise, c scores are centred around 0 for no bias; for item memory, a bias to responding ‘new’ would result in a positive score, while a bias to responding ‘old’ would result in a negative score. For source memory, a bias to responding ‘outer’ would result in a positive score, while a bias to responding ‘inner’ would result in a negative score. Data were aggregated by participant and emotional valence for the linear mixed model analyses. Signal detection theory measures were calculated using the psycho package in R (Makowski, 2018).

To calculate the effects of HP and emotional valence on each of these outcome measures, linear mixed models were compared. For each linear mixed model analysis, a null model specifying only a random effect of participant was first constructed. Then, models specifying each of the fixed (main) effects in turn was calculated, with a maximal model specifying the interactions between these main terms. Finally, an ANOVA was used to compare model performance, with the significance of main effects being inferred from the fixed effects structure(s) of the model or models which performed better than the null model. This procedure for model comparison uses Maximum Likelihood Estimates to balance model parsimony and variance explained (see e.g., Pinheiro & Bates, 2000). Visual inspection of plots comparing the fitted and residual values for each model did not reveal violations of the assumptions of linearity or homoscedasticity; likewise, an examination of qqplots for the residual values indicated no deviations from normality for each model. Linear mixed models were generated using the lme4 package (Bates et al., 2015); linear model scatterplots (Figures 6 and 7) were generated using the ggplot2 package (Wickham, 2016).

19 3. Results

Hallucination-proneness

Scores on the five auditory LSHS-R items ranged from 5 to 14 (out of a total possible range of 5 to 20), with an average of 9.13 (SD = 2.32). This result accords with the mean LSHS-R scores of 8.75 and 9.46 reported by Garrison and colleagues (2017) and McCarthy-Jones and Fernyhough (2011), respectively. Figure 2 displays the distribution of HP scores.

Accuracy

Accuracy rates were calculated for each participant, both overall and stratified by source (inner, outer, new) and emotional valence (positive, negative, neutral). Table 2 displays these aggregated accuracy scores. Average source accuracy indicates near chance performance. By contrast, average item accuracy for ‘new’ items indicates above chance performance.

Condition inner outer new

Valence overall .445 .423 .682

positive .503 .442 .409 .659

negative .500 .420 .436 .647

neutral .547 .473 .425 .742

Table 2: Source attribution accuracy rates by condition and emotional valence Figure 2: Frequency of hallucination-proneness scores

20 Discriminability and bias

3.3.1 Demographic effects

Before the effects of HP and emotional valence on d’ and c were calculated, ANOVAs examining the effects of age, education, and gender were conducted on both of these measures, for both item and source memory. These analyses revealed a significant effect of age on both item d’ (F(4,130) = 4.64, p = .002) and source d’ (F(4,130) = 4.93, p < .001). Figures 3 and 4 show item and source d’ by each of the age categories. The item d’ means for each age category were 1.20, 1.35, 0.85, 0.65, and 1.13, respectively; the source d’ means were 1.00, 0.86, 0.34, 0.37, and 0.37, respectively. Due to the presence of five age categories with unequal sample sizes, no post-hoc pairwise comparisons of age groups were conducted. The effect of age on these two measures is likely partially attributable to the relatively low number of people in the age bins 36-45, 46-55, and 56-65 (n of 5, 4, and 8, respectively). Nevertheless, because of the significant effect of age on d’ for both item and source memory, age was included as a fixed effect in the linear mixed models for item and source d’.

The analysis of item d’ also revealed a significant effect of gender (F(2,130) = 5.79, p = .004); non-binary people had a higher mean d’ (M = 2.06) than both males (M = 1.08) and females (M = 1.19). However, this effect is also likely an artefact of sample size, namely the small size of the non-binary group (n = 2), and indeed a t-test comparing item d’ between males and females yielded a non-significant result (t(62.92)= -0.90, p = .371). None of the demographic variables had a significant effect on c for either item or source memory.

Figure 3: Item discriminability plotted by age category

Figure 4: Source discriminability plotted by age category

21 3.3.2 Item memory

Unaggregated by emotional valence, average item d’ was 1.11 (SD = 0.48), and all participants had an item d’ greater than 0, indicating above chance discrimination between old and new items. Accordingly, item d’ was significantly greater than 0 (t(45)= 15.83, p < .001). Linear mixed model comparisons indicated no significant effect of HP score on item d’; however, a model specifying emotional valence as a fixed effect performed significantly better than a null model with only a fixed effect of age and a random effect of participant (χ2(1) =20.22, p < .001). Post-hoc Tukey’s t-tests revealed that neutral items (M = 1.44) were significantly more discriminable than both negative (M = 1.03; t(84) = 4.30, p < .001) and positive (M = 1.06; t(84) = 3.97, p < .001) items, but that there was no significant difference between the negative and positive items’ discriminability (t(84) = -0.34, p = .746). Figure 5 shows item d’ as it varies by emotional valence.

Unaggregated by emotional valence, average item c was less than 0.001 (SD = 0.42), and a t-test indicated this was not significantly different from 0 (t(45) < 0.001, p = .999). Linear mixed models indicated no significant effect of either HP score or emotional valence on item c (χ2(1) = 2.38, p = .123; χ2(1) = 0.42, p = .518, respectively).

3.3.3 Source memory

Unaggregated by emotional valence, average source d’ was 0.71 (SD = 0.57), indicating low ability to discriminate between inner and outer source; nevertheless, source d’ across all participants was significantly greater than 0 (t(45) = 8.41, p < .001). However, three

22

participants had an average source d’ of below 0, indicating below chance performance for discriminating between the sources; these participants were excluded from subsequent analyses. Linear mixed models yielded no evidence for an effect of either HP score or emotional valence on source d’. Figure 6 shows source discriminability (d’) plotted as a function of emotional valence and HP score.

Unaggregated by emotional valence, average source c was -0.16 (SD = 0.31), and a t-test indicated this was significantly less than 0 (t(45) = -3.42, p = .001), indicating a bias in the direction of reporting items as internal. One further participant was excluded from source analyses, for an outlying average source c of more than three standard deviations above the mean (M = 1.52). Linear mixed models revealed no effect of HP score on source c, but a model specifying valence group as a fixed effect generated significantly improved performance compared to a null model (χ2(1) = 5.04, p = .025). Post-hoc Tukey’s t-tests revealed the mean c of the negative items (M = -0.10) was significantly greater than that of the neutral items (M = -0.25; t(84) = -2.17, p = .033), and greater than positive items (M = -0.22) with marginal significance (t(84) = 1.74, p = .086). There was no difference between the positive and neutral c values (t(84) = -0.44, p = .665). Figure 7 shows source bias (c) plotted as a function of emotional valence and HP score, while Figure 8 shows average c by emotional valence.

23

While the linearity of the relationship between the outcome variables and fixed effects was confirmed by visual inspection of the fitted and residual values of each model, it is possible that the application of linear mixed models obscured the presence of a bi-directional bias whereby some participants displayed an internalising bias, and others an externalising bias. However, a normality test for source c (W = .98, p = .125) indicated no deviation from normality.

Figure 8: Source bias as a function of emotional valence Figure 7: Source bias as a function of HP and emotional valence

24 4. Discussion

In this experiment, the relationship between hallucination-proneness and verbal spatial source monitoring in healthy adults was tested. Overall, the results of this experiment do not provide support for the application of the source monitoring account to verbal spatial source monitoring in healthy adults. Regarding the first research question, HP score in the auditory modality did not have a significant relationship with spatial source discriminability or bias. There was no evidence for an increased spatial externalisation bias amongst high-HP adults as predicted by the source monitoring account, nor was there evidence for a bi-directional bias as reported by Stephane and colleagues (2010) in their sample of PwS-AVH+. These results thus align with those of Garrison and colleagues (2017) and those of McKague and colleagues (2012), who also found no evidence for a relationship between HP and spatial source discriminability or bias in healthy adults. By contrast, previous studies have found a relationship between HP in healthy adults and verbal source monitoring along the agentic source dimension (Larøi et al., 2004). The results of the current study thus indicate that, unlike agentic source, spatial source may not be a salient cognitive dimension of subclinical AVH-like experiences in the healthy population.

Because a spatial source monitoring bias is associated with AVH presentation amongst people with psychotic disorders (e.g., Franck et al., 2000), the lack of a relationship between HP and a verbal spatial source monitoring bias in healthy adults may have clinical significance. As discussed in the Introduction, studying HP in healthy adults has the potential to reveal which characteristics of AVHs can be used to distinguish clinical and subclinical AVHs. Given that some subclinical psychotic experiences develop into psychosis (van Os et al., 2009), the presence of a spatial source monitoring bias in a hallucination-prone healthy adult could potentially be useful as a clinical predictor of the development of psychosis. Thus, the continuum account of psychosis is not supported here due to the lack of evidence for a verbal spatial source monitoring bias, a result which has potential clinical significance for the identification of at-risk individuals with subclinical AVH-like experiences.

Not only was there a lack of a relationship between HP and an externalisation bias, as would have been predicted by the source monitoring account, but there was also a significant spatial source bias across participants towards reporting items as internal. This result accords with that of McKague and colleagues (2012) who also found a significant internalisation bias amongst low- and high-HP individuals (for positive items only, in the latter group). The appearance of an internalisation bias across two different experimental paradigms – that is, the

25

paradigm of McKague and colleagues who manipulated the perceived spatial source of other-generated speech, and the paradigm used here, in which the spatial source of self-other-generated speech was manipulated – implicates the presence of an internalisation bias along the spatial source dimension of speech in healthy adults irrespective of its agentic source. The cause of this internalisation bias cannot be extrapolated from the present data, and hence necessitates further investigation. It is possible, for example, that this spatial source internalisation bias represents an aspect of speech processing that is not captured by auditory HP but rather by some other aspect of psycholinguistic processing, such as the properties of inner speech. This possibility could be investigated by examining the relationship between verbal spatial source monitoring performance and the phenomenological characteristics of inner speech, as measured by the Varieties of Inner Speech Questionnaire (McCarthy-Jones & Fernyhough, 2011). Alternatively, the internalisation bias could represent a domain-general cognitive bias, which can be tested by investigating source attributions in other modalities, such as the visual modality. A final possibility for further research is to investigate the relationship between a verbal spatial source internalisation bias and other schizotypal traits such as delusion-proneness. In this way, the manifestation of an internalisation bias within verbal spatial source monitoring amongst healthy adults has the potential to reveal the differential cognitive mechanisms underlying this task amongst healthy adults compared to PwS-AVH+, which in turn may be used to identify the pathological impairments to verbal source monitoring in schizophrenia.

The analysis of the effects of emotional valence on verbal spatial source monitoring provided partial support for the expectations of research question 2. That is, the overall spatial source internalisation bias in this experiment was moderated by emotional valence, such that negative material had a significantly more positive (i.e., less internalising) bias than either neutral or positive material. Interestingly, McKague and colleagues (2012) report a similar result in their experiment, namely a marginally significant effect of emotional valence whereby negative material was less subject to an internalising bias than positive material. Spatial source monitoring in healthy adults thus appears to be moderated by emotional valence such that negative material is differentially subject to a spatial source monitoring bias. However, given the overall internalising bias across emotional valences, it cannot be extrapolated from this result that negative material was more subject to an externalising bias than other material, but rather only that negative material was less subject to an internalising bias than either neutral or positive material. It is therefore not possible to parse this result in terms of previous studies of the moderation of source monitoring by emotional valence, as those results have yielded an

26

externalisation bias with a greater externalisation effect for negative material. Furthermore, the effect of emotional valence on spatial source monitoring bias did not vary by HP score. The continuum model of psychosis is thus not supported by this result, as the emotional moderation of the agentic source monitoring bias found in PwS-AVH+ (e.g., Costafreda et al., 2008) was not found here in higher-HP adults.

Nevertheless, in addition to the effect of emotional valence on source bias as discussed above, emotional valence also had an effect on item discriminability, such that the discriminability between old and new items was significantly greater for neutral items than for either negative or positive items. This indicates that emotional valence, whether positive or negative, may have disrupted the regular binding processes which occur during memory encoding, and hence led to a decreased ability to distinguish whether these items had been seen before. This explanation would accord with previous research which has demonstrated the effect of emotional valence on memory accuracy, a process which may be disrupted in schizophrenia (see Dieleman & Röder, 2013). Thus, in conjunction with the effect of emotional valence on source bias such that negative items are less subject to an internalising bias, the results for item discriminability indicate that emotional valence does indeed have an impact on both recognition and source memory encoding processes in healthy adults.

There are two important factors which constrain the interpretability of the results discussed above. Firstly, the recall task appears to have been difficult for participants. Overall, source memory, the measure most crucial to the research hypotheses, was low; while source monitoring discriminability was significantly above chance, as indicated by the d’ score, it did not generate a large score range, which may have decreased the likelihood of finding an effect of HP score or emotional valence on source memory discriminability and bias. By contrast, while item memory was good, as indicated by moderate d’ for item memory and high overall ‘new’ item accuracy, it did not reach a ceiling effect, indicating that participants faced some difficulties even in recognising which items they had seen before. This effect may be attributable to a high item number; item number is a trade-off in any experimental methodology, as lower item number compromises analytic and statistical power. This low power is particularly an issue within the signal detection theory analysis employed here, wherein outcome measures are calculated by aggregating by both participant and condition. In addition to high item number, accuracy may also have been influenced by the overall low concreteness of the adjectives; this is a trade-off with the use of personal adjectives, which will be discussed in 4.2 Future Research. Nevertheless, to increase variability in source monitoring scores, future studies may recalibrate item number and item concreteness. Another way to

27

increase source monitoring accuracy would be to tell participants that they will be later asked to recall source (e.g., McKague et al., 2012). However, in a paradigm which employs self-generated speech, this methodology can lead to participants employing idiosyncratic, construct-irrelevant strategies to increase their memory scores, for example producing all externally-generated words in a memorable tone of voice. Hence, this measure was not taken here, but may be appropriate for other methodologies.

Secondly, the lack of a relationship between HP and either measure of spatial source monitoring may be in part attributable to sample size, more specifically the relative lack of variability across the HP spectrum generated within the sample. That is, across a possible range of 5 to 20 for the auditory HP score, the maximum score was 14. Further research is therefore required to confirm the lack of a relationship between HP and verbal spatial source monitoring across a larger variability in HP scores.

Limitations

Three aspects of the methodology limit the generalisability of the findings. Firstly, a quantitative approach was taken in this study to categorise the words by valence (see section 2.2.1 Source monitoring stimuli). However, employing semantic rating tasks in which participants are asked not only for a quantitative rating of valence but also for a categorical judgment might yield a more representative measure of valence category. Secondly, the influence of the acoustic properties of each individual’s inner speech on their performance in the source monitoring task was not examined. Future studies may investigate whether the acoustic properties of the individual’s inner reading voice (e.g., Vilhauer, 2017) influence the discriminability between internally- and externally-generated speech. The relationship between the acoustic properties of inner reading voice and HP could also be potentially informative, for example, if individuals with higher HP are more likely to have an inner voice with a non-self agentic source. Such a result would have implications for the application of the source monitoring account to hallucination-prone individuals. Finally, no measure of either state or trait emotional characteristics was taken here; future studies should adopt such a measure, as both types of emotional characteristic have been linked to both HP and source monitoring abilities (see e.g., Kanemoto et al., 2013).

Future Research

These results indicate that spatial source monitoring in healthy adults may be moderated by the emotional valence of verbal material. As discussed in the Introduction (see section 1.3), a prominent account of the role of emotional valence in verbal source monitoring attributes the greater externalisation bias for negative material to an anxiety-reduction mechanism, which

28

leads to negative thoughts – especially those about the self – being externalised as derogatory AVHs. An informative avenue for future research would therefore be to investigate the extent to which this effect of emotional valence is attributable to the ‘personal’ characteristic of the words used. That is, both personal and non-personal adjectives were employed here, and it was beyond the scope of this study to quantify the extent to which each adjective could be used as a personal descriptor. (For example, consider the words ‘blunt’ and ‘impossible’: these may carry considerably different emotional valences depending on whether they are applied to a person or an object.) While some researchers have circumnavigated this difficulty by exclusively employing unambiguous personal adjectives (e.g., Johns et al., 2006), the use of only adjectives which can be used as personal descriptors may be a trade-off with concreteness; as discussed in the Limitations section, the low concreteness of the adjectives employed in this study may have accounted for the low source monitoring accuracy. Hence, in future studies, researchers could employ a task wherein participants are asked to rate each adjective for the extent to which they believe that adjective applies to themselves (e.g., Morrison & Haddock, 1997). This task serves the dual purpose of increasing the cognitive effort applied to the processing of abstract personal adjectives and therefore their memorability, and would allow researchers to directly test whether an externalisation bias of negative material amongst PwS-AVH+ and/or high-HP individuals can be attributed to the material being applied to the self. The role of self-attribution of emotionally valent material could also be investigated by directly comparing source monitoring of personal and non-personal adjectives. It would be informative to firstly quantify the extent to which an adjective can be used as personal along both literal and metaphorical dimensions using a semantic rating task, as well as the relative frequency of each usage.

Future research may also expand the methodology used to collect source attributions. That is, the three-alternative decision (‘inner’, ‘outer’, and ‘new’) utilised here functionally resembles paradigms in which participants are only asked to make source attributions for ‘old’ items. However, this methodology does not facilitate a full examination of the relationship between item and source memory. Specifically, researchers have previously investigated whether there is a unified memory signal underlying both item and source memory by taking ratings and source attributions for both old and new items, and plotting the resulting receiver operating characteristic functions. Such research has shown that the accuracy of item memory and source memory attributions is related, but that the relationship between their confidence ratings may be curvilinear (see Starns, Rotello, & Hautus, 2014), and single- and dual-process signal detection models make diverging predictions with regard to the linearity of this

29

relationship. Neither ratings nor source attributions for new items were taken here; this decision was in part due to the high item number, and the fact that taking post-item ratings takes time and thus constrains item number. However, for more in-depth information concerning the underlying distribution of the memory signals which contribute to item and source attributions, future studies could include a source attribution decision for ‘new’ items as well as ‘old’ items, in addition to rating scales for each of these decisions. Such research would have the potential to reveal whether the relationship between item and source memory is disrupted in individuals with either a clinical or subclinical manifestation of AVHs. Employing this design in future research would also facilitate a comparison of overall accuracy scores between the different conditions, regardless of error type; this type of comparison is not possible under signal detection theory analyses, which only take into consideration the attributions made along one parameter (i.e., old-new, inner-outer) at a time.

Conclusions

In the present study, the application of the source monitoring account of AVHs to hallucination-proneness in healthy adults was tested by examining whether healthy adults with higher hallucination-proneness are more likely to make spatial source monitoring errors in a verbal source monitoring paradigm. The results yielded no evidence for a relationship between verbal spatial source monitoring and hallucination-proneness, although the difficulty of the task and the lack of variability in hallucination-proneness scores constrain the interpretability of this absent effect. Nevertheless, the presence of an overall internalisation bias for verbal material, as well as the moderation of this bias by negative material specifically, indicates that verbal spatial source monitoring in healthy adults may tap into similar psycholinguistic processes to those underlying clinical AVHs. Suggestions for future research have been made to further investigate the characteristics of verbal spatial source monitoring in healthy adults. By comparing the source monitoring deficits of clinical and non-clinical AVH-hearers with such research, we should improve our understanding of the cognitive and psycholinguistic deficits which differentiate clinical and non-clinical manifestations of AVHs. This understanding can in turn be used to inform therapeutic practice for the identification and management of clinically significant characteristics of AVHs in psychosis.

30 References

Alba-Ferrara, L., De Erausquin, G. A., Hirnstein, M., Weis, S., & Hausmann, M. (2013). Emotional prosody modulates attention in schizophrenia patients with hallucinations. Frontiers in Human Neuroscience, 7, 59.

Alderson-Day, B., Smailes, D., Moffatt, J., Mitrenga, K., Moseley, P., & Fernyhough, C. (2019). Intentional inhibition but not source memory is related to hallucination-proneness and intrusive thoughts in a university sample. Cortex, 113, 267–278.

Allen, P., Amaro, E., Fu, C. H., Williams, S. C., Brammer, M. J., Johns, L. C., & McGuire, P. K. (2007). Neural correlates of the misattribution of speech in schizophrenia. The British Journal of Psychiatry, 190, 162–169.

American Psychiatric Association (2013). Diagnostic and statistical manual of mental disorders (5th ed.). Washington, DC: Publisher.

Anselmetti, S., Cavallaro, R., Bechi, M., Angelone, S. M., Ermoli, E., Cocchi, F., & Smeraldi, E. (2007). Psychopathological and neuropsychological correlates of source monitoring impairment in schizophrenia. Psychiatry Research, 150, 51–59.

Baker, C. A., & Morrison, A. P. (1998). Cognitive processes in auditory hallucinations: attributional biases and metacognition. Psychological Medicine, 28, 1199–1208. Bates, D., Maechler, M., Bolker, B., & Walker, S. (2015). Fitting Linear Mixed-Effects Models

Using lme4. Journal of Statistical Software, 67, 1–48.

Beavan, V., Read, J., & Cartwright, C. (2011). The prevalence of voice‐hearers in the general population: A literature review. Journal of Mental Health, 20, 281–292.

Bendall, S., Jackson, H. J., & Hulbert, C. A. (2011). What self-generated speech is externally misattributed in psychosis? Testing three cognitive models in a first-episode sample. Schizophrenia Research, 129, 36–41.

Bentall, R. P., & Slade, P. D. (1985a). Reality testing and auditory hallucinations: a signal detection analysis. British Journal of Clinical Psychology, 24, 159–169.

Bentall, R. P., & Slade, P. D. (1985b). Reliability of a scale measuring dispositions towards hallucinations: A brief report. Personality and Individual Differences, 6, 527–529. Brébion, G., Amador, X., David, A., Malaspina, D., Sharif, Z., & Gorman, J. (2000). Positive

symptomatology and source-monitoring failure in schizophrenia - an analysis of symptom specific effects. Psychiatry Research, 95, 119–131.

Brookwell, M. L., Bentall, R. P., & Varese, F. (2013). Externalizing biases and hallucinations in source-monitoring, self-monitoring and signal detection studies: A meta-analytic review. Psychological Medicine, 43, 2465–2475.