Introduction
Auditory verbal hallucination or voice hearing in the absence of a corresponding external stimuli, referred to as “voices” henceforth, is understood to be on a continuum (Linscott & van Os, Reference Linscott and van Os2010). It has a lifetime prevalence rate of up to one in ten individuals in the general population (Maijer et al., Reference Maijer, Begemann, Palmen, Leucht and Sommer2018) associated with reduced distress (Taylor & Murray, Reference Taylor and Murray2012), and it occurs in a range of mental health disorders including psychosis, where voices are usually reported as more severe and distressing and linked with increased clinical needs (Larøi, Reference Larøi2012) and need for care (Johns et al., Reference Johns, Kompus, Connell, Humpston, Lincoln, Longden and Larøi2014). Voice distress in psychosis has been linked with increased anxiety, depression, and a higher degree of negative voice content (Scott, Rossell, Toh, & Thomas, Reference Scott, Rossell, Toh and Thomas2020b), and the presence of posttraumatic stress disorder (PTSD; de Bont et al., Reference de Bont, van den Berg, van der Vleugel, de Roos, de Jongh, van der Gaag and van Minnen2015).
Voices in psychosis have been associated with a history of different types of trauma (Bailey et al., Reference Bailey, Alvarez-Jimenez, Garcia-Sanchez, Hulbert, Barlow and Bendall2018; Grindey & Bradshaw, Reference Grindey and Bradshaw2022), with voice content at times directly reflecting trauma content (e.g., hearing the voice of a past abuser; van den Berg et al., Reference van den Berg, Tolmeijer, Jongeneel, Staring, Palstra, van der Gaag and Hardy2023). Traumatic stress-induced changes (i.e., traumagenic neurodevelopmental model; Pruessner et al., Reference Pruessner, Cullen, Aas and Walker2017) are implicated in emotional memories being encoded without contextual information (Brewin & Burgess, Reference Brewin and Burgess2014). Trauma memories might be experienced on a continuum of contextualized autobiographical memories and fragmented sensory experiences, such as voices, that are appraised as externally sourced when re-experienced (Hardy, Reference Hardy2017).
Disrupted care experiences during childhood, including neglect, abuse, and early losses, are involved in the development of fearful attachment (van Ijzendoorn et al., Reference van Ijzendoorn, Schuengel and Bakermans–Kranenburg1999), one of three types of insecure attachment patterns developed during early caregiving relationships (Ainsworth & Bell, Reference Ainsworth, Bell and Steinberg1981). A higher prevalence of fearful attachment has been observed in individuals with psychosis (Carr et al., Reference Carr, Hardy and Fornells-Ambrojo2018a), with significantly higher levels of hallucinations and voice distress (Bucci et al., Reference Bucci, Emsley and Berry2017).
Cognitive models of psychosis propose that childhood relational trauma and insecure attachment styles renders individuals more vulnerable to negative interpretations of self and others (Garety et al., Reference Garety, Kuipers, Fowler, Freeman and Bebbington2001; Scott, Rossell, Meyer, et al., Reference Scott, Rossell, Meyer, Toh and Thomas2020a). These interpretations, especially when originating from experiences of subordination and marginalization, influence hearers’ appraisals of voices as more malevolent, persecuting, powerful, critical and are related to greater voice distress (Birchwood et al., Reference Birchwood, Meaden, Trower, Gilbert and Plaistow2000; Larøi et al., Reference Larøi, Thomas, Aleman, Fernyhough, Wilkinson, Deamer and McCarthy-Jones2019). In non-clinical hearers, voices are reported as predominantly benevolent and less distressing (Daalman et al., Reference Daalman, Boks, Diederen, de Weijer, Blom, Kahn and Sommer2011).
Relational theories also implicate interpersonal frameworks and social world relationships in how the hearer responds to voices, influencing distress (Birchwood et al., Reference Birchwood, Gilbert, Gilbert, Trower, Meaden, Hay and Miles2004; Thomas et al., Reference Thomas, McLeod and Brewin2009). It is then important to consider social factors and their effect on voices experiences (e.g., culture-wide/culture−specific views of voices; Luhrmann et al., Reference Luhrmann, Padmavati, Tharoor and Osei2015), accounting also for how social factors’ associations to voices are attenuated by age, gender, and ethnicity (Bonoldi et al., Reference Bonoldi, Simeone, Rocchetti, Codjoe, Rossi, Gambi and Fusar-Poli2013). Thus, relational therapies aim to target distress by focusing on such relationships, enabling the individual to state their needs and gain awareness (Craig et al., Reference Craig, Rus-Calafell, Ward, Leff, Huckvale, Howarth and Garety2018; Hayward et al., Reference Hayward, Berry and Ashton2011).
Most research to date has investigated the interplay between trauma, attachment, and voice appraisals in isolation or paired associations, with debates over the specificity (i.e., one process leads to voices) and equifinality (i.e., voices may occur via different processes from a variety of different initial conditions) of the multiple mechanisms and factors involved (Gibson et al., Reference Gibson, Alloy and Ellman2016). Such traditional variable-centered studies provide estimated parameters indicating how factors are related in all individuals, which are assumed to be drawn from a single population, without considering how such factors may have different interactions which could differ across subpopulations of individuals (Morin et al., Reference Morin, Boudrias, Marsh, McInerney, Dagenais-Desmarais, Madore and Litalien2017). Conversely, person-centered analyses are data-driven (e.g., Begemann et al., Reference Begemann, Sommer, Brand, Oomen, Jongeneel, Berkhout and Bell2022), relaxing this assumption, and aimed to unearth multiple specific combinations of factors in cross-sectional data based on how they differ or are similar in traits and dimensions of interest (Saunders et al., Reference Saunders, Buckman and Pilling2020).
This study aimed to (1) use latent profile analysis to identify statistically distinct groups of individuals who hear distressing voices, based on the interplay between experiences of trauma, fearful attachment style and voice appraisals, and (2) explore whether identified profiles are differentially associated with demographic characteristics and clinical outcomes. Involvement of Experts by Experience was incorporated at all stages of the analysis, in line with recommendations in research (i.e., Corstens et al., Reference Corstens, Longden, McCarthy-Jones, Waddingham and Thomas2014; National Institute for Health and Care Research [NIHR], 2012) and with the aim that chosen measures and interpretation of findings had more meaningful and real-world clinical applications.
Methods
Setting
The baseline assessment data from the AVATAR2 multisite (London, Manchester, Glasgow) randomized controlled trial (RCT; see Garety et al., Reference Garety, Edwards, Ward, Emsley, Huckvale, McCrone and Craig2021 for detailed description) testing the efficacy of two versions of AVATAR therapy (Craig et al., Reference Craig, Rus-Calafell, Ward, Leff, Huckvale, Howarth and Garety2018; Leff et al., Reference Leff, Williams, Huckvale, Arbuthnot and Leff2014) against treatment-as-usual was used for this study. The study was approved by London-Camberwell St Giles Research Ethics Committee (REC/HRA ethical approval 20/LO/0657) which included consent for participation in ancillary/additional studies that utilize anonymized data collected, such as the current study. The recruitment, data collection of assessment meetings, and data storage were carried out according to the AVATAR2 trial protocol (please see Garety et al., Reference Garety, Edwards, Ward, Emsley, Huckvale, McCrone and Craig2021, Reference Garety, Edwards, Jafari, Emsley, Huckvale, Rus-Calafell and Ward2024 for further information). The baseline assessment meetings completed asked participants about their experiences of their voices, overall mood, and wellbeing via multiple measures.
Participants
The 345 participants included in this study met the inclusion and exclusion criteria of the AVATAR2 RCT (for full details, see Garety et al. (Reference Garety, Edwards, Ward, Emsley, Huckvale, McCrone and Craig2021) and Supplementary Material A).
Measures
Table 1 displays measures, their descriptions, and psychometric information, for the indicator variables included in the LPA, and the distal variables in post-LPA analyses, which are demographic characteristics and clinical presentation outcomes. For further details on measures and participant characteristics, see Table 1 and Supplementary Material B.
Table 1. Table with indicator and distal variables included in the latent profile analysis and further analyses and their measure description and properties (see Supplementary Material B for further details)

Analysis
Latent profile analysis
LPA is a mixture modeling approach extending latent class analysis (Vermunt & Magidson, Reference Vermunt, Magidson, Hagenaars and McCutcheon2002) to include both continuous and categorical variables (Gibson, Reference Gibson1959). LPA was conducted using Mplus version 8 (Muthén & Muthén, Reference Muthén and Muthén2017) evaluating whether there are distinct groups (latent profiles; LPs) of voice hearing individuals based on the observed individual response patterns on the indicator variables of appraisal of voices (as malevolent, omnipotent, benevolent, and voice engagement or resistance), the presence of fearful attachment style, traumatic experiences, and association of trauma-voice content. Further information can be found in Supplementary Material C about the selection of indicator variables and negotiating with model convergence constraints.
Model fit of the LPA models was compared using The Vuong-Lo-Medell-Rubin likelihood ratio test (VLMR-LRT; Lo et al., Reference Lo, Mendell and Rubin2001), and Bootstrap likelihood ratio difference test (B-LRT; Nylund et al., Reference Nylund, Asparouhov and Muthén2007) alongside Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and entropy-based criterion values (Geiser, Reference Geiser2013). FIML is used for systematically missing data in LPA. The full information about the model selection process can be found in Supplementary Material D.
Associations between profiles and both demographic information and clinical outcomes
Following the LPA, a series of analyses were performed to explore the associations between a) LPs and demographic variables (age, gender, ethnicity) and b) LPs and clinical presentation outcomes (voice severity, PTSD symptoms, level of global distress and motivation and pleasure difficulties).
To account for misclassification accuracy in LPA (Bakk et al., Reference Bakk, Tekle and Vermunt2013; Clark & Muthén, Reference Clark and Muthén2009), multinomial logistic regression analyses via the bias-adjusted R3STEP method (Asparouhov & Muthén, Reference Asparouhov and Muthén2014) was used, with the reference profile for associations in a) being the largest sample size. Specifically, the recommended Bolck-Croon-Hagenaars (BCH) procedure (Bolck et al., Reference Bolck, Croon and Hagenaars2004) was employed, with demographic variables (age, gender, ethnicity) included as confounders in models in b) examining differences in clinical outcomes across profiles (McLarnon & O’Neill, Reference McLarnon and O’Neill2018; Nylund-Gibson et al., Reference Nylund-Gibson, Grimm and Masyn2019). The profile-specific regression intercepts yielded were compared in omnibus Wald Chi-square tests and pairwise z-tests that suggest how belonging to one profile is differentially associated with a clinical outcome beyond the demographics and other profiles (Clark & Muthén, Reference Clark and Muthén2009). Missing data on demographic and clinical variables were handled using multiple imputation in Mplus (Section 11.1 of Asparouhov & Muthén, Reference Asparouhov and Muthén2021). For detailed information about these post-LPA analyses, see Supplementary Material E.
Patient and public involvement
In line with the Hearing Voices Movement (Corstens et al., Reference Corstens, Longden, McCarthy-Jones, Waddingham and Thomas2014) and the NIHR (2012) recommendations to involve people with lived experiences in research, patient and public involvement (PPI) has played a key role at all stages for co-production of the AVATAR2 trial, including design, recruitment of staff and participants, data collection, analysis, and dissemination. An active and creative group of people was established, comprising 20 members across all four sites, from diverse backgrounds, with lived experience of psychosis and recovery, and including carers. PPI co-production was continued in this study for decision-making processes and to ensure that interpretations of the complex findings were clinically useful and with real-world reflections.
First, one-to-one consultations were conducted with five PPI consultants at the pre-analysis stage. They provided feedback about the importance of the chosen indicator variables, the retention of measures (e.g., trauma-voice content association (i.e., TVAQ; Woods et al., Reference Woods, Jones, Alderson-Day, Callard and Fernyhough2015), and the voice appraisals measure (i.e., BAVQ-R (Chadwick et al., Reference Chadwick, Lees and Birchwood2000) and not the Voices Acceptance and Action Scale (Shawyer et al., Reference Shawyer, Ratcliff, Mackinnon, Farhall, Hayes and Copolov2007), or Voice Power Differential Scale (Birchwood et al., Reference Birchwood, Meaden, Trower, Gilbert and Plaistow2000) and informed the meaningful inclusion of ethnicity and gender demographics (i.e., exploring their association with the LPA model).
Consultations were also completed post-LPA-analysis with four PPI consultants. This explored their insights and understanding of the identified LPA groupings and their indicator variables’ distributions. A qualitative analysis of the rich feedback was beyond the scope of this study; thus, PPI reflections were incorporated when discussing the results. For further detailed information and summary of consultations, see Supplementary Material F.
Results
Descriptive statistics
The full sample of 345 participants had an average age of 39.6 years old, with a higher percentage of male gender (61.4%), ‘White’ ethnicity (59.7%), followed by ‘Black or mixed Black’ ethnicity (20.9%), and the majority had a Schizophrenia or Schizoaffective diagnosis (87.2%). The full sample results, from the RSQ and Mini-TALE, found a high rate of reported interpersonal adversities of fearful attachment and trauma (only 5.8% reported no trauma) alongside negative voice appraisals (i.e., omnipotence, malevolence) and resistance to voices, with lower occurrence of benevolent voices and engagement. Such indicator variables were mostly weakly to moderately associated. As expected, participants had high voice severity and emotional distress, high levels of motivation and pleasure difficulties (albeit just below cutoff score indicative of negative symptoms; Li et al., Reference Li, Li, Zou, Yang, Xie, Yang and Chan2018), and a PTSD dimensional score indicative of a middle level of symptom severity (M = 10.6 out of a total of 24). See full details of the full sample statistics in Supplementary Material G.
Latent profile analysis
The VLMRT-LRT and B-LRT yielded significant p-values (p < 0.05), alongside decreases in the AIC and BIC values, when comparing successive models from a two-profile to a four-profile solution. At the five-profile model, the VLMR-LRT p-value increased to above 0.05 (p = 0.198), and the entropy value for the four-profile versus the five-profile model was slightly higher, therefore the four-profile solution was considered optimal for the data (see Supplementary Material D for model comparison statistics).
Latent profiles descriptions
Descriptive statistics for the full sample and distributions for each latent profile (LP) are displayed in Table 2 (presented graphically in Figure 4 in Supplementary Material H), with the description of each profile provided here:
-
1. LP1 (largest; approximately 44.5%, n = 157) is described as ‘Adverse voices and relational trauma’ – In comparison with other profiles, individuals in this profile have the highest scores, which are higher as compared to the full sample, of fearful attachment style, trauma, and are more likely to believe that trauma and voices are related. Compared to the full sample, they report moderately higher on beliefs of voices being malevolent, omnipotent, and resisting voices, the highest scores comparative to other profiles. It has the lowest scores of benevolent voice appraisal and engagement, akin to LP3.
-
2. LP2 (second largest; approximately 25.2%, n = 84) is described as ‘Low malevolent and omnipotent voices’– Individuals in this profile are set apart by their scoring the lowest on all subscales relating to negative voice appraisals of omnipotence, malevolence and resistance to voices, compared to other profiles and the full sample. Compared to the full sample, individuals in this group have a marginally lower number of traumas experienced, traumas’ relatedness to voices and fearful attachment. Benevolent voice appraisal and engagement scores were low, similar to the full sample, however not as low as LP1 and LP3.
-
3. LP3 (second smallest; approximately 16.5%, n = 57) is described as ‘Adverse voices yet low relational trauma’ – Individuals in this profile are reporting the lowest scores across profiles for relational traumas. Compared to the full sample they have much lower experiences of fearful attachment, moderately a lower number of traumas reported, and where such traumas are not believed to be associated to voices heard. Although, such individuals are also scoring the second highest on subscales relating to omnipotent and malevolent voice appraisals, which is marginally higher than average full sample. Resistance to voices is similar to the high score in the full sample. Additionally, akin to LP1, they have a lower score of benevolent voice appraisals and engagement.
-
4. LP4 (smallest; approximately 13.8%, n = 47) is described as ‘High benevolent voices’ – Individuals in this profile are set apart by their scoring very high as compared to the full sample and other profiles on benevolent voice appraisal and engagement with voices. Although, they are similar to the full sample, with a higher number of traumas, beliefs of voices being related to traumas and experiencing fearful attachment and higher scores of omnipotent voice appraisals. Scores of malevolence voice appraisals and resistance to voices were low, lower than the full sample scores.
Table 2. Full sample and latent profiles indicator variable distribution

a see Supplementary Material B for full measures information.
Exploring profiles’ association with demographic information and clinical presentation outcomes
The next step was to analyse whether (a) LPs identified were associated with demographics (age, gender, ethnicity) and (b) LPs were differentially associated with outcome variables which represented participants’ clinical presentations (severity of voices, PTSD symptoms, global distress and motivation and pleasure difficulties). See Supplementary Material I for full statistical results of these analyses, with summaries presented later.
a) Demographic factors associated with profiles
The descriptive statistics for demographics are shown in Table 3, where there is a higher proportion of male gender across all profiles. LP1 ‘Adverse voices and relational trauma’ has the highest proportions of female and other non-male genders as compared to other profiles. The multinomial logistic regressions considering the association of demographics and LPs showed a significant association between individuals being female and other non-male genders and an increased likelihood of belonging to LP1 ‘Adverse voices and relational trauma’ as compared to LP2 ‘Low malevolent and omnipotent voices’ (OR [95% CI] = 0.472 [0.247; 0.901]). For ethnicity, only individuals within the ‘Any other’ ethnicity category, which has the lowest percentage, as compared to ‘White’, were significantly more likely to belong to LP4 ‘High benevolent voices’ as compared to LP1 ‘Adverse voices and relational trauma’ (reference profile; OR [95% CI] = 8.78 [2.75; 28.03]).
Table 3. Distribution of demographic covariates across profiles and full sample

a Please note that when considering statistical testing of the demographic covariate associations, due to sample size (i.e., only 4 other non-male respondents), female and other non-male gender categories were collapsed together, thus the analysis includes a binary variable of ‘male’ and ‘female and other non-male genders’.
b Please see Table 2 in Supplementary Material B for full details of ethnicity for each of the categories in line with the larger AVATAR2 RCT.
b) Clinical presentation outcomes associated with profiles
Table 4 shows the distribution of clinical outcomes in the full sample and across LPs. The profile-specific intercept regressions showed that all profiles have a unique influence on the distal outcomes independent of the influence of the demographic covariates accounted for in the model (treated as binary: male/female and other non-male genders), and white/not-white (i.e., all other ethnicities)). When comparing these profile-specific intercepts in equivalence omnibus chi-square tests, these indicated that LPs were only significantly different in their associations with voice severity (X 2(3) = 10.4, p = 0.015) and global distress outcomes (depression: X 2(3) = 8.4, p = 0.038*; anxiety: X 2(3) = 10.2, p = 0.017; stress: X 2(3) = 10.3, p = 0.016), yet not with PTSD (X 2(3) = 7.2, p = 0.066) and motivation and pleasure difficulties (X 2(3) = 6.7, p = 0.083) outcomes, after controlling for covariates’ influences.
Table 4. Latent profiles and associated clinical outcome distal variables

a see further information about measures in Supplementary Material B.
b DASS-21 is scored via doubling summed scores for each subscale given it is the short form for the scale.
c Missing cases are represented here for the descriptive statistics however multiple imputation was utilized for missing values on distal outcomes for the statistical analyses.
Thus, considering the pairwise z-tests for the significant outcomes, LP1 ‘Adverse voices and relational trauma’ closely followed by LP3 ‘Adverse voices yet low relational trauma’ (not statistically different from each other) had the highest voice severity scores, relative to the full sample, which were both statistically different in their influence of such outcomes when compared to LP2 ‘Low malevolent and omnipotent voices’ and LP4 ‘High benevolent voices’ (not significantly different between themselves) which were associated with lower voice severity scores. Additionally, LP1 ‘Adverse voices and relational trauma’ was significantly associated with individuals having the worst emotional distress (ranges: severe depression and stress, extremely severe anxiety) as compared to better outcomes in LP2 ‘Low malevolent and omnipotent voices’ (ranges: moderate depression and anxiety, mild stress). Specifically for depression, LP1 ‘Adverse voices and relational trauma’ and LP4 ‘High benevolent voices’ (second highest mean score of severe depression) are significantly different in their association with depression outcomes. See Figure 5 in Supplementary Material I for a graphical representation of these results.
Table 5 provides a visual summary of all findings.
Table 5. Summary of Latent Profile results and clinical presentation outcome with traffic light system representing severity*

Note: The color scheme represents a red-amber-green traffic light system to visually indicate severity of the clinical presentation outcomes across profiles (red: highest severity to green: lowest severity, with amber in between).
Discussion
This study identified four statistically different subgroups of participants based on their experiences of fearful attachment, trauma, beliefs about trauma-voice relatedness, and voice appraisals: LP1 ‘Adverse voices and relational trauma’, LP2 ‘Low malevolent and omnipotent voices’, LP3 ‘Adverse voices yet low relational trauma’, and LP4 ‘High benevolent voices’. Differential profile membership was significantly associated with outcomes of voice severity and global distress but not for other clinical outcomes (internal experience of motivation and pleasure and related behaviors and PTSD). LP1 and LP2 reflect the two ends of the severity spectrum of adversities and voice relationships theorized in various models of voices, whereas the latter two revealed profiles (LP3, LP4) suggest a novel intersection of voice experiences and adversity. There were no clear patterns of demographic effects across profiles, with only gender association with LP1 and no differences across the more represented minority groups of Black and Asian backgrounds. The reflections from PPI consultations are incorporated throughout the discussion of results.
Interpersonal adversity and impact on voices: support for existing models
In line with the existing models associating trauma and insecure attachment with voice appraisals and severity (e.g., Bailey et al., Reference Bailey, Alvarez-Jimenez, Garcia-Sanchez, Hulbert, Barlow and Bendall2018; Berry et al., Reference Berry, Varese and Bucci2017; Scott, Rossell, Meyer, et al., Reference Scott, Rossell, Meyer, Toh and Thomas2020a), the larger LP1 ‘Adverse voices and relational trauma’ had higher interpersonal adversities linked to negative voice appraisals with worst voice severity and emotional distress outcomes. LP2 ‘Low malevolent and omnipotent voices’ had better scores on clinical outcomes alongside less adverse factors linked to substantially lower negative voice appraisals. PPI consultants discussed how the experience of cumulative trauma has a role in voice appraisals of powerlessness and persecution, in line with relational models which posit that interpersonal adversities can be internalized, individuals see themselves lower in power, mirrored in their voice relating (Hayward, Reference Hayward2003; Thomas et al., Reference Thomas, McLeod and Brewin2009) and more distressed responses to voices (Pilton et al., Reference Pilton, Bucci, McManus, Hayward, Emsley and Berry2016). In this profile, trauma memories or dissociative trauma mechanisms, which can be predisposed by fearful attachment (Puckett et al., Reference Puckett, Sood and Newman-Taylor2023), could result in trauma-related memories encoded in fragmented ways, re-experienced as negative voices (Hardy, Reference Hardy2017; Pilton et al., Reference Pilton, Varese, Berry and Bucci2015).
The likelihood of belonging to the first more severe profile was associated with individuals being more likely to be female or other non-male genders as compared to the second profile. Increased severity and voice hearing distress has been reported in females in some studies (Murphy et al., Reference Murphy, Shevlin, Adamson and Houston2010; Suessenbacher-Kessler et al., Reference Suessenbacher-Kessler, Gmeiner, Diendorfer, Schrank, Unger and Amering2021; Toh et al., Reference Toh, Gurvich, Thomas, Tan, Neill and Van Rheenen2020), compared to males (see Barajas, Ochoa, Obiols, & Lalucat-Jo, Reference Barajas, Ochoa, Obiols and Lalucat-Jo2015 for less favorable outcomes in males), with passive relating to voices explaining this association (Schlier et al., Reference Schlier, Sitara, Strauss, Rammou, Lincoln and Hayward2021) and male gender norms promoting the minimization of reporting voices (Goldstein & Lewine, Reference Goldstein, Lewine, Castle, McGrath and Kulkarni2000), powerlessness, or distress (Parent, Hammer, Bradstreet, Schwartz, & Jobe, Reference Parent, Hammer, Bradstreet, Schwartz and Jobe2018).
Unexpected adversity and voice profiles
LP3 ‘Adverse voices yet low relational trauma’ and LP4 ‘High benevolent voices’ had an interesting combination of psychosocial factors, which may have been overlooked in the literature perhaps given previous studies’ different methodologies (e.g., isolated/paired psychosocial associations; Saunders et al., Reference Saunders, Buckman and Pilling2020). PPI consultants suggested that, where much lower fearful attachment and traumas in the LP3 group were reported, other ongoing environmental adversities (e.g., urbanicity, living in deprived areas) or difficult experiences (e.g., being from a minority group and experiencing subordination, discrimination, bullying), that were not captured in our demographic variables, can be impacting the individual. Where a continuing sense of threat results in hypervigilance which has been implicated in making negative information more salient shaping more negative voice-content (Larøi et al., Reference Larøi, Thomas, Aleman, Fernyhough, Wilkinson, Deamer and McCarthy-Jones2019). As such, reflected in LP3 voices being highly malign and omnipotent, with the second-worst voice severity outcome.
Alternative non-trauma routes to voices in LP3 could be a genetic/family history (van Winkel et al., Reference van Winkel, Van Nierop, Myin-Germeys and van Os2013), cognitive vulnerability, such as, strategies using punishment and worry for controlling unwanted thoughts, which are implicated with more psychological dysfunction and distress (Morrison & Wells, Reference Morrison and Wells2000), and low self-esteem (Williams et al., Reference Williams, Bucci, Berry and Varese2018). This group’s lower emotional distress could be linked with the lower fearful attachment suggesting more robust attachment styles (e.g., secure) may contribute to this group’s better emotion regulation and coping (e.g., lower emotional hyperactivity linked to positive psychosis symptom vulnerability; Berry et al., Reference Berry, Varese and Bucci2017). It may also be indicative of aspects not captured in this study, such as voice acceptance lowering negative distress from the lesser emotional resistance to voices, impacting voice distress (Varese et al., Reference Varese, Morrison, Beck, Heffernan, Law and Bentall2016) and psychological flexibility’s positive impact on general emotional wellbeing (Morris et al., Reference Morris, Garety and Peters2014).
LP4 ‘High benevolent voices’ has a substantially higher level of benevolent voice appraisals, highly engaged with, alongside reported high fearful attachment, trauma, and omnipotent voices, similar to the first profile. With less supportive attachment relationships to provide a buffer to traumas reported, PPI consultants and previous literature suggest that other experiences from an upbringing in cultures and religious contexts, with helpful positive beliefs in relation to voices, could provide a soothing or accepting stance for voice hearing (Larøi et al., Reference Larøi, Luhrmann, Bell, Christian, Deshpande, Fernyhough and Woods2014). Voices then can be experienced as powerful, as seen in this group, yet this power is reported as more benevolent (e.g., Cottam et al., Reference Cottam, Paul, Doughty, Carpenter, Al-Mousawi, Karvounis and Done2011). Of note, being of ‘Any other’ (and not other core minority Black and Asian backgrounds) rather than ‘White’ ethnicity was associated with this profile (as compared to LP1) and further information would be needed to interpret this finding.
In line with other studies showing a positive relationship between benevolent voices and lower voice distress (Sanjuan et al., Reference Sanjuan, Gonzalez, Aguilar, Leal and Van Os2004), the greater voice engagement may help lower the intensity and severity of voices in LP4 (Sayer et al., Reference Sayer, Ritter and Gournay2000), potentially due to a voice dialogue where the hearer is alongside the voice, rather than a victim, as suggested by PPIs. However, when considering LP4’s greater emotional distress, in line with psychological flexibility models, an appraisal of voices as omnipotent in this profile (regardless of it being a benevolent intention) can imply greater judgment toward these experiences and be responded to literally in subordinate ways which can impact emotional wellbeing (Gilbert et al., Reference Gilbert, Birchwood, Gilbert, Trower, Hay, Murray and Miles2001). Additionally, experiencing voices as friendly in this profile, especially in the context of potentially impoverished social interactions, may be associated with retreating to the voice relationship for comfort/companionship (Miller et al., Reference Miller, O’Connor and DiPasquale1993), impacting social functioning and seeking treatment support (Favrod et al., Reference Favrod, Grasset, Spreng, Grossenbacher and Hodé2004). Of importance, the current study’s participants are seeking treatment of a relational kind. Yet, the social withdrawal (LP4’s average motivation and pleasure difficulties being close to cutoff indicative of negative symptoms) may have associations with higher chronicity of symptoms, previously proposed to be that with more turbulent patients they are prone to more positive relationship with voices, and poorer outcomes including emotional distress (Favrod et al., Reference Favrod, Grasset, Spreng, Grossenbacher and Hodé2004).
Limitations
This is a cross-sectional study where LPA identifies subgroups dependent on the variables included which was limited to the available sample size (Tein et al., Reference Tein, Coxe and Cham2013; see Supplementary Material C). LPA is recognized as not sufficient to prove that profiles found will exist as tangible groups of people in other data samples, they provide guides for clinical consideration, but are not suggested as reified profiles (Williams & Kibowski, Reference Williams, Kibowski, Jason and Glenwick2016). It thus should be noted that this study sample is specific individuals distressed by persistent voices who view a relational AVATAR treatment approach as relevant to them which can influence findings, since other voice hearers might not conceptualize their voice experience as a relationship (e.g., Chin et al., Reference Chin, Hayward and Drinnan2009). Thus, this is not representative of all voice hearers, some of whom might not engage with services.
The trauma measure (i.e., TALE) used, although a clinically useful measure, requires further better-quality evidence (Airey et al., Reference Airey, Taylor, Vikram and Berry2023) which incurs limitations for the reliability of interpreting trauma associations in this study. The selection of the shorter (Mini-TALE) measure in the AVATAR2 trial assessment protocol (Garety et al., Reference Garety, Edwards, Ward, Emsley, Huckvale, McCrone and Craig2021, Reference Garety, Edwards, Jafari, Emsley, Huckvale, Rus-Calafell and Ward2024) was intended to minimize participant burden from using longer intrusive questionnaires (Fornells-Ambrojo et al., Reference Fornells-Ambrojo, Johns, Onwumere, Garety, Milosh, Iredale and Jolley2017). This measure captures only the cumulative elements of trauma, without differentiating childhood and adulthood trauma. Given evidence about the specific role of childhood trauma (Stanton et al., Reference Stanton, Denietolis, Goodwin and Dvir2020) and the cumulative or shared impact of adversity across different life stages (Pastore et al., Reference Pastore, de Girolamo, Tafuri, Tomasicchio and Margari2020; Trauelsen et al., Reference Trauelsen, Bendall, Jansen, Nielsen, Pedersen, Trier and Simonsen2015) in psychosis, the findings of the current study require replication with a more robust trauma measure that distinguishes trauma across the lifespan and includes other trauma types, which PPI feedback for all LPs highlighted as important, such as, poverty, discrimination, and urbanicity factors (Fett et al., Reference Fett, Lemmers-Jansen and Krabbendam2019; Varchmin et al., Reference Varchmin, Montag, Treusch, Kaminski and Heinz2021). There are inconsistent findings related to associations of trauma type and hallucination modality (c.f. Barnes et al., Reference Barnes, Emsley, Garety and Hardy2023). Thus, due to statistical constraints related to model convergence (see Supplementary Materials C for further consideration), this study was not able to separate the types of interpersonal traumas (c.f. Begemann et al., Reference Begemann, Sommer, Brand, Oomen, Jongeneel, Berkhout and Bell2022) since this would complicate the model and increase error where with the current sample size this would yield meaningless profiles. Non-trauma-related factors such as comorbid difficulties (e.g., anxiety and depression, autism spectrum disorder) could also have a mediating role for all profiles, perhaps impacting cognitive functioning and schemas of self-other, as such, consideration of such further factors are important and suggested by PPIs in future studies.
Interpretability of variables in a clinical utility sense were continually negotiated, however, for analysis ethnicity categories needed to be collapsed into mixed groups and binary codes making findings harder to interpret (i.e., association of ‘Any other’ ethnicity and increased likelihood of membership to LP4 ‘High benevolent voice’). This highlights the limitations in investigating the role of ethnicity within studies not specifically powered for this, as is the case of the current study.
Clinical and research implications
The current findings highlight the importance of thorough assessments of adverse experiences, alongside careful formulations of the meaning traumas have in reference to voices and how they relate to ways the voice hearer does/does not and has/has not been able to form safe attachments with others. Given profiles from a sample of individuals with distressing voices have both negative and positive voices, asking patients for detailed descriptions and interpretations of their voice could aid clinicians to not miss information, such as, benevolent appraisals and engagement with voices valued by individuals and where fewer interpersonal adversities co-occur with distressing negative appraisals. Further exploration of other factors with stronger influences, including other adversities and comorbid difficulties not measured in this study, should be explored.
These case formulations could support clinicians and service users to discuss appropriate tailored treatment. For example, where relational therapies address past relationships with abusive caregivers via how these are represented in the voice interaction, aiming to change relating behavior with voices (O’Brien et al., Reference O’Brien, Rus-Calafell, Craig, Garety, Ward, Lister and Fornells-Ambrojo2021). Additionally, incorporating what the individual values about benevolent voices, especially respecting cultural or religious aspects, can help interventions to shape the voice relationship to how the individual wants, so they can feel safer and are meeting their own needs.
The findings in this study would benefit from further research considering generalizability of current profiles in other psychosis populations and non-clinical voice hearers, since this is a sample of individuals distressed by persistent voices who were willing to participate in the AVATAR2 RCT. A longitudinal investigation of identified profiles and their association with outcomes would be informative. Such as, AVATAR2 studies considering whether identified profiles respond differently to the AVATAR treatments received, which may inform understanding of how best to target treatments to different groups. The invaluable collaboration from PPIs in this study, in shaping variable selection and meaning making of results, is an important demonstration of how Experts by Experience can and should be incorporated in future research so that this can refine psychological support to meet the needs of voice hearers. Details of cultural background, historical wellbeing prior to voices, relationship to help, protective family experiences, occupational context, and developmental stages (stress/pressure) were raised by PPIs which should be incorporated in further research.
Conclusion
This study uses LPA to provide new insights related to the complex interplay of interpersonal adversities co-occurring with positive and negative voice appraisals that are differentially associated with voice and emotional distress. Clinical information from this LPA can inform individualized assessments including careful consideration of voice appraisals, especially benevolent voices, and how these are linked to interpersonal adversities, to support decisions around helpful interventions. Information from the identified profiles could inform services, audits, and evaluations.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S003329172500008X.
Acknowledgments
The research was funded by the Wellcome Trust (FWBC-AVATAR 098272/z/12/z). The views expressed in this research are those of the author(s) and not necessarily those of the NHS, the NIHR or the Department of Health.
Funding statement
This study is funded as part of the AVATAR2 Randomized Controlled Trial which is funded by The Wellcome Trust Ltd., through an Innovations Project award (grant reference 215471/Z/19/Z). The funding body has no role in the design of the study or the collection, analysis and interpretation of data or the writing of the manuscript. This research was part of JM’s Doctorate in Clinical Psychology thesis which received funding from the UCL Research Department of Clinical, Educational and Health Psychology.
Competing interest
All authors have no conflicts of interest or funding.
Ethical standard
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.