Evidence for Reduced Sensory Precision and Increased Reliance on Priors in Hallucination-Prone Individuals in a General Population Sample

Abstract

Background

There is increasing evidence that people with hallucinations overweight perceptual beliefs relative to incoming sensory evidence. Past work demonstrating prior overweighting has used simple, nonlinguistic stimuli. However, auditory hallucinations in psychosis are often complex and linguistic. There may be an interaction between the type of auditory information being processed and its perceived quality in engendering hallucinations.

Study Design

We administered a linguistic version of the conditioned hallucinations (CH) task to an online sample of 88 general population participants. Metrics related to hallucination-proneness, hallucination severity, stimulus thresholds, and stimulus detection rates were collected. Data were used to fit parameters of a Hierarchical Gaussian Filter (HGF) model of perceptual inference to determine how latent perceptual states influenced task behavior.

Study Results

Replicating past results, higher CH rates were observed both in those with recent hallucinatory experiences as well as participants with high hallucination-proneness; CH rates were positively correlated with increased prior weighting; and increased prior weighting was related to hallucination severity. Unlike past results, participants with recent hallucinatory experiences as well as those with higher hallucination-proneness had higher stimulus thresholds, lower sensitivity to stimuli presented at the highest threshold, and had lower response confidence, consistent with lower precision of sensory evidence.

Conclusions

We replicate the finding that increased CH rates and recent hallucinations correlate with increased prior weighting using a linguistic version of the CH task. Results support a role for reduced sensory precision in the interplay between prior weighting and hallucination-proneness.

Introduction

Computational techniques have increasingly been used in recent years to characterize cognitive and perceptual mechanisms that may underlie psychotic symptoms, with the goal of facilitating the development of novel treatments.^1–8 One finding that has gained prominence is the importance of the overweighting of prior beliefs in the generation of positive symptoms, such as hallucinations.^1,2,5,8–15 Previous work demonstrates that perceptual systems, rather than relying directly upon input from the sensory organs, use sensory input to update probabilistic models of the causes of sensations (the environment).^11,16,17 In this view, perception is an inferential process in which organisms infer what is around them by combining sensory input with their prior beliefs about the world, weighted by the reliability of these different information channels. This weighted blending of sensory input and priors can be observed in many common situations (eg, the use of lip-reading cues¹⁸ and sentence context¹⁹ in understanding speech in noisy environments; the use of shading to infer 3-dimensional surfaces in the visual domain²⁰). Bayesian statistical models have been used to investigate this combination of sensory input and prior knowledge¹⁷; these models have succeeded in predicting performance on a number of perceptual tasks^21–23 as well as neural activity in sensory contexts.²⁴ Substantial recent work has established that, at least in some cases, overreliance on priors compared with incoming sensory evidence may underpin hallucinations.⁵ While evidence has pointed to this overweighting of priors, what is less clear is precisely how this takes place: is the relative overweighting of priors seen in hallucinations independent of or related to the absolute weighting (and quality) of incoming sensory information?^1,2,8,25

This question is by no means purely academic; understanding how prior overweighting and resultant hallucinations relate to upstream sensory processing errors could inform different interventions based on identification of different distal mechanisms. For example, in silico simulations^1,2 demonstrated that an overweighting of certain priors can generate hallucinations, though this only occurs in the setting of permissively low sensory precision. Benrimoh et al^1,2 theorized that maladaptive priors would be encoded in upper levels of the processing hierarchy with weighting of some priors modified by dopaminergic signaling; whereas lower-quality sensory evidence could be due to reduced integrity of white matter connections such as the arcuate fasciculus (which is often degraded in schizophrenia²⁶) or encoded by changes in cholinergic tone, also known to be altered in schizophrenia.²⁷ This view of maladaptive priors being expressed as perceptual changes only when sensory precision is permissively low suggests a different mechanism from, eg, a scenario where there is no absolute change in prior weighting but only low weighting of sensory evidence arising from low-quality sensory information, resulting in a relative overweighting of priors, used to fill in missing sensory data.

Here, we attempt to interrogate this question of relative prior and sensory weighting using the conditioned hallucinations (CH) task.^5,12,13 CH occur as a result of classical conditioning, where a subject is presented with a salient stimulus paired with a difficult-to-detect target (eg, an image and a sound) at the same time in a repeated manner, such that in the presence of the salient stimulus (eg, the image) and the absence of the target, the subject may hallucinate the target in order to satisfy their expectation of the stimulus being present. Our group has developed and validated several versions of this task, which has been shown to be sensitive to susceptibility to psychosis and hallucinations^5,12 and current hallucination state.¹³ However, this task has not yet been administered in a large, general population sample, which would allow for a nuanced assessment of the interplay between sensory processing and hallucination-proneness outside of frank psychopathology or the presence of frank hallucinations.

Methods

Participants

Participants were recruited and completed the experimental task using Amazon Mechanical Turk. In order to be included in the analysis, participants needed to report the absence of a current or past psychotic illness. It is important to note that this was not a healthy control sample, but rather a nonpsychotic general population sample. This was necessary because removing participants with any mental illness may have artificially reduced the number of participants who demonstrate hallucination-proneness, given that there is some evidence of more frequent psychopathology (such as anxious and depressive symptoms) in those with hallucination- or delusion-proneness.^28–30

One hundred and twenty participants signed up to complete the task. Of these, 21 did not complete the task. Eight completed the task but provided data that did not pass quality control screening.¹³ Two subjects endorsed having a medical condition that could cause hallucinations and their data were discarded. These participants were removed because we wished to evaluate performance on the CH task in a general population sample of those without a psychotic disorder (regardless of cause), which would hold relevance to the utility of this task for screening and prediction of outcomes in a similar sample. Similarly, 1 subject endorsed drug or alcohol use in the 2–3 hours preceding their participation and their data were discarded. This resulted in 88 subjects available for analysis. Within these 88 subjects, one reported being red-green color blind and 3 reported corrected-to-normal (rather than normal) hearing; post hoc analysis revealed no significant differences between their median CH rates and the rest of the sample (H₁ = 0.75, P = .39), and, as such, these subjects were included in final analysis.

Measures

Participants completed several questionnaires prior to the CH task. These included the Patient Health Questionnaire-9 (PHQ-9) and Generalized Anxiety Disorder-7 (GAD-7) to screen for depression and generalized anxiety symptoms, respectively^31,32; the Launay-Slade Hallucination Scale (LSHS)³³ to measure hallucination-proneness; the Auditory Hallucination Rating Scale (AHRS), to determine the intensity and nature of any recent hallucinatory experiences³⁴; the Adverse Childhood Events and Trauma History Questionnaire (THQ)³⁵ given the strong association between psychosis and trauma³⁶; as well as demographic information and psychiatric history. We chose to use the LSHS because it is a commonly used measure of hallucination-proneness in the general population, and principal component analyses have previously shown that it measures factors related to vivid mental events, hallucinations with a religious theme, and auditory and visual hallucinatory experiences.³⁷ The AHRS was designed to track hallucination severity over a 24-hour period, originally in the context of studies using transcranial magnetic stimulation to treat auditory hallucinations in psychotic individuals.³⁸

For analysis, the sample was split into groups based on hallucination-propensity, which was generally low (mean LSHS score: 7.1 out of 60; SD: 9.4; see table 1). Participants were designated as having high hallucination-propensity if they scored at least 1 SD above the mean on the LSHS (n = 15); all other participants were designated as low propensity (n = 73). We chose 1 SD to create groups with meaningfully different symptom profiles. Because this was a general population sample, LSHS scores were generally low, such that a median split would have created a high-propensity group that mixed both participants with low and high scores. We also conducted all analyses using the continuous questionnaire scores, allowing us to determine whether the continuous results mirrored binarized group results. The AHRS, which focuses more on recent hallucinatory experiences, had a mean of 4.8 (SD = 7.2) out of 49 possible points, again demonstrating a low rate of hallucinatory experiences in this population. We also split the sample into subgroups based on recent hallucination status using their responses to the AHRS. Those with an AHRS score of 0 (ie, with no significant recent hallucinatory experiences) were put into the “no recent hallucinations” subgroup (AH−), and those with a score above 0 were put into the “some recent hallucinatory experiences” subgroup (AH+); each subgroup had 44 members.

Table 1.

Demographic and Clinical Information by Group

	Group by AHRS Score		Group by LSHS Score
	AH−	AH+	Low Propensity	High Propensity
n (%)	45 (51.1)	43 (48.9)	73 (83.0)	15 (17.0)
Age, mean (SD)	42.0 (10.8)	36.1 (10.1)	40.9 (10.9)	30.5 (4.3)
Sex, n (%M)	29 (64.4)	33 (76.7)	48 (65.7)	14 (93.3)
Race, n (%)
American Indian/Alaska Native	0 (0.0)	3 (6.9)	2 (2.7)	1 (6.7)
Asian	5 (11.1)	5 (11.6)	8 (11.0)	2 (13.3)
Black or African American	1 (2.2)	4 (9.3)	4 (5.5)	1 (6.7)
White	36 (80.0)	31 (72.1)	57 (78.1)	10 (66.7)
More than 1 race	3 (6.7)	0 (0.0)	2 (2.7)	1 (6.7)
Mental illness, n (%)	4 (8.9)	1 (2.3)	5 (6.8)	0 (0.0)
Medication current, n (%)	1 (2.2)	2 (4.7)	2 (2.7)	1 (6.6)
LSHS, mean (SD)	2.2 (3.7)	12.3 (10.8)	3.3 (4.0)	25.8 (3.9)
AHRS, mean (SD)	0.0 (0.0)	9.8 (7.5)	2.5 (4.2)	16.0 (8.3)
ACE, mean (SD)	1.2 (1.9)	0.93 (1.3)	1.2 (1.7)	0.7 (1.2)
THQ, mean (SD)	2.2 (2.1)	1.1 (1.6)	1.9 (2.0)	0.20 (0.77)
PHQ-9, mean (SD)	2.4 (3.2)	6.4 (6.6)	3.0 (4.3)	10.8 (6.2)
GAD-7, mean (SD)	1.6 (2.5)	4.0 (5.0)	2.0 (3.5)	6.3 (4.9)
Peters Delusion Inventory (PDI) total, mean (SD)	1.6 (2.4)	1.5 (2.0)	1.4 (2.0)	2.3 (3.3)

Note: AHRS, Auditory Hallucination Rating Scale; LSHS, Launay-Slade Hallucination Scale; THQ, Trauma History Questionnaire. Bolded values indicate significant difference (P < .05) between AHRS and LSHS subgroups.

Materials

Participants used their own computers to complete the CH task (see below), and did so in their environment of choice. All participants were asked to use headphones, ensure that their keyboard was working, and keep their screen brightness and volume at a maximal level. Prior work has demonstrated that the combination of these hardware checks and thresholding (which uses participant responses rather than hard-wired stimulus intensities) results in insensitivity of task performance to hardware, operating system, and browser differences.¹³ All stimuli included in the task were presented via React (https://reactjs.org/). Participants completed the questionnaires prior to engaging in the CH task, using a standard HIPAA-compliant questionnaire data collection interface (REDCap @ Yale).

Conditioned Hallucination (CH) Task

Each participant completed the 3 different versions of the CH task in random order. This was done in order to gather pilot data for an affective manipulation (showing images prior to the task), which is not the focus of this analysis. Here, we focus on only the “neutral” condition, where subjects saw only gray squares before the task started.

The linguistic CH task was adapted from past instantiations of the original CH task^5,12,13 which uses Pavlovian associated-learning to elicit hallucinations (figure 1a). In the CH task, participants are instructed to report if they detected an auditory target embedded in 70 db-SPL white noise. The target is always paired with a salient visual cue. In the original task, the auditory target was a 1-kHz pure tone. In this version, participants were instructed to detect a voice that said a nonsense word. A verbal target stimulus was chosen to more closely mimic hallucinations in psychosis, as has been done previously in general population samples.¹⁵

Fig. 1.

Conditioned hallucination (CH) task structure. (a) One of 3 visual patterns and voice pairs were presented simultaneously. White noise played throughout the task. After stimuli were presented, participants indicated by button-press whether they heard the voice or not and then rated confidence in their decision. (b) Using the QUEST thresholding procedure, we estimated 75%, 50%, and 25% detection thresholds for each participant. Presentation of stimuli at each intensity level was systematically varied over the course of 12 blocks of 15 trials each, with tone-trials becoming more infrequent and quieter, and absent-tone trials becoming more frequent.

Participants first completed a short practice to familiarize themselves with the CH task. Participants repeated the practice until they reached an accuracy of at least 70%, as in past work.^5,13 After this, individual threshold estimates were obtained using the maximum-likelihood-based QUEST procedure.^5,39 We then fit the 75% threshold to a standard psychometric curve to determine target volumes at which participants were likely to report detection at rates of 50% and 25%.

Consistent with prior work,^5,12,13 participants completed 12 blocks of the target detection task at each of the thresholds in addition to no-voice trials during which only the visual pattern was presented. As past work has demonstrated the emergence of group-wise difference within the first half of the experiment,¹³ we halved the number of trials per block from 30 to 15 to minimize fatigue. In order to smoothly build the association between the auditory target and the accompanying visual stimulus, the probability of at-threshold trials decreased over the blocks and subthreshold and target-absent trials increased (figure 1b).

Each version of the task had a different pattern-word pairing: red vertical stripes and “dob,” green horizontal stripes and “bez,” or black and white checkerboard and “yag.” This was done to avoid stimulus pairs learned in 1 version of the task influencing CH rates in a version of the task completed later. As in past work, CH were defined as the participant indicating the presence of a voice stimulus on a trial when no target was present (ie, the no-target trials).^5,12,13 The CH rate was then defined as the proportion of no-target trials in which the participant reported hearing a voice. We found no significant relationship between CH rates and stimuli used during the neutral condition or order of task versions (Stimuli: H₂ = 1.7, P = .43; Order: H₂ = 0.72, P = .69), QUEST-derived threshold (Stimuli: H₂ = 0.40, P = .82; Order: H₂ = 0.85, P = .65), AHRS subgroups (Stimuli: χ² = 0.85, P = .65; Order: χ² = 1.0, P = .60) or LSHS subgroups (Stimuli: χ² = 3.1, P = .21; Order: χ² = 0.36, P = .84).

In addition to reporting detection/nondetection of the target, participants were asked to hold down the response button to indicate their degree of confidence in this judgment, guided by a visual scale appearing at fixation and moving from 1 (Unsure) to 5 (Confident) as the response button was held down.

Hierarchical Gaussian Filter Analysis

To pinpoint potential mechanisms driving behavioral differences associated with hallucination-proneness, we employed the Hierarchical Gaussian Filter (HGF) model, a hierarchical Bayesian model of learning in a dynamic environment (see figure 4a).^40,41 A 3-tiered version of the HGF has been adapted to use trial-wise response and stimulus intensity values from the CH task^5,12,13 Extensive details of this model have been previously published.^5,12,13 The model estimates belief states during the task. Level 1 (μ₁) corresponds to the belief that there was a voice during a given presentation of the visual stimulus. Level 2 (μ₂) corresponds to the belief that, in general, a voice is predicted by the presence of the visual stimulus. Level 3 (μ₃) corresponds to belief in the volatility of the association between voice and the visual stimulus (ie, beliefs about the volatility of the relationship captured by level 2). Critically, the model also allows the estimation of parameters reflecting the weight afforded to (ie, the precision of) prior expectations (ν), evolution rates in the strength of association (ω₂), and volatility (ω₃) beliefs. Finally, the model translates posterior beliefs in the presence of the tone on any given trial into a probability of reporting detection using a softmax function, the slope of which corresponds to inverse decision noise (β⁻¹). To implement the HGF, we used the TAPAS computational toolbox (github.com/translationalneuromodeling/tapas) which is freely available with all relevant code.

Fig. 4.

High prior precision and decision noise are related to hallucination severity on Hierarchical Gaussian Filter (HGF) analysis. (a) The HGF model uses stimulus strength (U) and responses to estimate latent states driving behavior on the linguistic conditioned hallucination task. These states include: belief trajectories at 3 levels (X₁, X₂, and X₃), the evolution rates of these beliefs (ω and θ), relative precision of prior (ν), and decision noise (β⁻¹). X₁ corresponds with trial-wise probability that a voice was present. X₂ reflects the global probability that there is a voice when the pattern is present. X₃ is belief in the volatility in the association between the voice and pattern. For those who had experienced hallucinations within the last 24 hours (b–d), we did not find significant differences in belief states at X₃ (b, top) or X₂ (b, middle). AH+ had stronger trial-wise beliefs that a voice was present (b, bottom) and weighted priors more heavily (c). However AH− and AH+ groups did not differ in decision noise (β⁻¹; d). For groups reflecting general hallucination severity (e–g). They did not significantly differ in belief states (e) nor how heavily they weighted their priors (f). However, they did have higher decision noise (g). * P < .05; ** P < .01

Previous work using the HGF with an online version of the CH task¹³ used the empirical grand mean detection rates (eg, the actual mean detection rates at each threshold) instead of expected detection rates (eg, 25%, 50%, and 75%) as the intensities for each condition when fitting the model. This was intended to account for discrepancies between intended detection threshold and the actual, ground-truth detection rates observed. We explored which version of the model best explained the data. We performed Bayesian model comparison using spm_BMS.⁴² Model exceedance probabilities showed support for the expected means (PXP = 0.514) over the empirical (PXP = 0.486). Additionally, a paired t test of the inversion results from each model indicated a significant increase in percent-identical responses for the expected (90.6%) over the empirical (90.4%; t₈₇ = 2.74, P = 7.30 × 10⁻³). Thus, we performed all analyses using the expected means as the ground-truths.

Statistical Analyses

We aimed to determine if performance on the CH task was sensitive to hallucinations proneness in this general population sample. Thus, we used regression to test the relationship between CH rates and total score of the LSHS and AHRS. We also tested for mean differences between low- and high-proneness groups, and those reporting vs not reporting recent hallucinatory experiences, as outlined below. Prior to analysis, participant response data were used to fit the parameters of a 3-tier HGF, described in detail in prior work.^5,12,13 This was done in order to estimate latent states driving behavior and behavioral differences between groups in the task. In past work, we have demonstrated that the parameter ν, which represents the weighting of prior beliefs relative to that of sensory evidence during perception, as particularly important for conditioned and clinical hallucination-propensity.^5,12,13 Exploratory analyses sought to compare groups on secondary measures of task performance and HGF parameter estimates. Based on a power analysis using initial pilot data, a sample size of 90 was determined to be appropriate to detect differences in computational parameters derived from participant responses. This power analysis was specifically conducted with the intent to discover differences in the v parameter between groups based on symptom level. Pilot data were not included in this analysis.

Behavioral data were not normally distributed, being skewed toward a low CH rate; this was expected in a general population sample. As such, nonparametric statistical tests were used where possible. Kruskal-Wallace tests were used to examine differences between participant subgroups for each condition, and linear models were used to determine the relationship between hallucination-proneness and parameters of interest. Repeated measures ANOVA⁴³ was used to assess relationships between subgroups over trials and blocks, as there is no equivalent nonparametric test for 3-way comparisons. The primary purpose of this analysis was to ensure that this linguistic version of the CH task was capable of replicating our previous findings that hallucination-proneness and severity were related to higher CH rates and prior hyper-precision. Thus, group-wise tests of mean CH rates and prior precision were our primary analyses. Analyses utilizing the continuous versions of scales were secondary analyses, intended to add detail to the group-wise analyses. All other analyses were exploratory in nature. We did not correct for multiple comparisons given the exploratory nature of this analysis. Analyses were carried out using R Version 4.1.0⁴⁴ and Python Version 3.9.13.

In addition, exploratory analyses of metacognition were carried out (see supplementary material).

Results

Sample Characteristics

Table 1 displays demographic and clinical characteristics of participants retained for analysis separated by hallucination-proneness and recent hallucinatory experiences subgroups. Groups did not differ by sex or race, with the population primarily composed of individuals identifying as male (n = 62; 70.5%) and Caucasian (n = 67; 76.1%). The mean age of the participants retained for analysis was 39.5 (SD: 11.2), with a range of 20–66. The high hallucination-proneness groups were significantly younger than the low-proneness groups (AH− and AH+: t₈₆ = −2.64, P = .01; Low- and High-Propensity: t₅₆ = −6.18, P = 7.60 × 10⁻⁸). 90% of the sample were participants based in the United States, and all participants reported English as their primary spoken language. One participant endorsed regular cannabis use in the last 4 weeks, and 2 endorsed higher-than-recommended levels of alcohol consumption. No participants reported currently being under the influence of a substance. A total of 5 subjects endorsed a history of mental illness, including depression, social anxiety, and OCD (supplementary table S1); 1 participant endorsed current antidepressant use but no other participants endorsed active use of medications.

While the majority of the sample did not endorse depression or anxiety, 18.2% did endorse clinically significant depressive symptoms (≥ moderate symptom severity on PHQ-9), and 11.3% endorsed significant anxiety (≥ moderate symptom severity on GAD-7). With respect to trauma, the mean number of traumatic events experienced was 1.6 (SD: 1.9) and the mean number of adverse events experienced specifically in childhood (ACE score) was 1.1 (SD: 1.7), with 9 participants (10.2%) reporting an Adverse Childhood Experiences (ACE) score of 4 or more.

Hallucination-proneness subgroups defined by both hallucination measures (LSHS and AHRS) did not differ in the proportion reporting psychiatric diagnosis or prescribed medications, but did differ in the general severity and presence of depression and anxiety symptoms. High-proneness groups endorsed higher clinically relevant depressive (AH− and AH+: t₆₀ = 3.62, P = 6.03 × 10⁻³; Low- and High-Propensity: t₁₇ = 4.63, P = 2.44 × 10⁻³) and anxiety (AH− and AH+: t₆₂ = 2.8, P = 6.47 × 10⁻³; LSHS: t₁₇ = 3.22, P = 4.98 × 10⁻³) symptoms. Conversely, low-proneness groups endorsed greater recent trauma (AH− and AH+: t₈₀; Low- and High-Propensity: t₅₇ = −5.68, P = 4.69 × 10⁻⁷). Groups did not differ in either childhood trauma or delusional ideation. Together, results are consistent with previous studies of psychosis-proneness.^29,45

Hallucination-Proneness Is Associated With Poorer Sensory Performance and Greater CH Rates

High-proneness groups had higher QUEST-derived thresholds than low-proneness groups (AH− and AH+: H₁ = 5.76, P = .016; Low- and High-Propensity: H₁ = 8.61, P = 3.33 × 10⁻³; figures 2a and 2e), meaning that they required higher stimulus intensity levels in order to report detection. This relationship tracked with severity of both recent hallucinatory experience and hallucination-proneness (AHRS total score: F_1,86 = 13.9, P = 3.42 × 10⁻⁴, R = 0.36; LSHS total score: F_1,86 = 12.7, P = 6.00 × 10⁻⁴, R = 0.39; supplementary figures S1a and S1c).

Fig. 2.

Hallucination-proneness is associated with poorer sensory performance and greater CH rates. (a, e) Estimated 75% detection thresholds. Estimated thresholds were higher for those who reported any hallucinations in the last 24 hours (AH+, a) and those with high general propensity for hallucinations (High-Propensity, e). (b, f) Percent of reported voice-detected trials at each estimated threshold. (c, g) Percent of No-Voice trials reported as voice present (CH Rate). (d, h) Mean confidence for each response by target volume and response. Black boxes indicate significant differences. In general, hallucination-prone groups were less confident in their decisions; however, confidence was not significantly different for any given condition between AH− and AH+ groups. High-proneness groups were significantly less confident when reporting no on No-Voice, 25%, and 50% and when reporting yes on 75% conditions. *P < .05; ***P < .001. Note: CH, conditioned hallucination.

As observed in previous work,^5,13 our primary analysis demonstrates that those with high hallucination-proneness were more likely to experience CH (μ_AH− = 8.8% [12.0], median: 5.6%, μ_AH+ = 21.6% [25.4], median: 8.5%; AH− and AH+: H₁ = 5.67, P = .017; μ_Low = 13.2% [19.8], median = 5.6%, μ_High = 23.4% [23.3], median = 9.9%; Low- and High-Propensity H₁ = 6.16, P = .013; figures 2c and 2g). Also consistent with recent findings, in secondary analyses we show that these differences tracked more with severity of recent hallucinations (F_1,86 = 8.37, P = 4.83 × 10⁻³, R = 0.28; supplementary figure S1b) than with general propensity (F_1,86 = 3.86, P = .05, R = 0.18; supplementary figure S1d).

Because of observed differences in QUEST-derived thresholds between proneness groups, we performed exploratory analyses to investigate the relationship between QUEST-derived detection thresholds and task performance. We did not find that threshold value itself was predictive of detection during 75% condition trials (F_1,86 = 2.19, P = .14) or CH rate (F_1,86 = 0.24, P = .63). However, we did find that, despite having higher threshold volumes, those with high general proneness were less likely to report detecting the tone on 75% conditions (H₁ = 4.93, P = .026; figure 2b). We observed a significant negative relationship between detection at the 75% condition with both general (F_1,86 = 4.00, P = .049; supplementary figure S1d) and recent (F_1,86 = 6.02, P = .016; supplementary figure S1b) hallucination-proneness.

Hallucination-Prone Groups Are Less Confident in Reporting Detection on Target-Present Trials and Have Reduced Confidence When Reporting Nondetection in Target-Absent Trials

While AH+ and AH− groups did not differ in overall confidence levels (figures 2d and 2h), in our exploratory modeling of confidence we found significant interactions between responses (yes or no) and group as well as a 3-way interaction between response, group, and condition (no voice, 25%, 50%, 75%; supplementary table S2). Post hoc analyses revealed that, as severity of recent hallucinations increased, confidence in detecting the target on the 75% condition (F_1,86 = 4.80, P = .03, R = −0.20) and when reporting nondetection on the no-target (F_1,86 = 9.19, P = 2.26 × 10⁻³, R = −0.30), 25% (F_1,86 = 6.34, P = .01, R = −0.24) and 50% (F_1,85 = 4.02, P = .048, R = −0.18) conditions decreased. We also explored this relationship using our general proneness metrics (Low- and High-Propensity groups, LSHS-total) using repeated measures ANOVA (subgroups; supplementary table S3). We observed similar effects as AHRS grouping; however, there was a main effect of LSHS group on confidence. Post hoc analyses indicated significantly decreased confidence in the high-severity group when reporting yes on the 75% condition (H₁ = 5.70, P = .02) and reporting nondetection on the no-target (H₁ = 6.32, P = .012), 25% (H₁ = 4.64, P = .03), and 50% (H₁ = 4.03, P = .047) conditions. These effects also scaled with severity of symptoms, as regression analyses indicated significant negative relationships between LSHS total score and confidence with each condition (No-Voice, no: F_1,86 = 17.4, P = 7.31 × 10⁻⁵, R = −0.40; 25%, no: F_1,86 = 11.93, P = 8.59 × 10⁻⁴, R = −0.33; 50%, no: F_1,85 = 10.83, P = 1.45 × 10⁻³, R = −0.32); 75%, yes: F_1,86 = 12.6, P = 6.29 × 10⁻⁴, R = −0.34).

High-Proneness and High-Severity Groups Maintain High CH Rates Over Time

We additionally explored how propensity for CH changed throughout the course of the task (figures 3a and 3e). As the experiment progressed, CH rates decreased and then stabilized. For the low-proneness groups, the drop occurred more rapidly and plateaued at a lower CH rate than in the high-proneness groups. A repeated measures ANOVA (supplementary tables S4 and S5) revealed a significant effect of group (AH− and AH+: F_1,86 = 11.68, P = 9.67 × 10⁻³; Low- and High-Propensity: F_1,86 = 6.92, P = .010) and a group-by-trial interaction (AH− and AH+: F_1,6158 = 8.20, P = 4.2 × 10⁻³; F_1,6158 = 61.3, P = 5.74 × 10⁻¹⁵).

Fig. 3.

High-proneness and high-severity groups maintain high CH rates over time. (a, e) Cumulative CH rates. Both AH+ and High-Propensity groups were more likely to report hearing a voice in No-Voice trials. There was also a significant interaction with block, indicating that groups differed in learning rates. This trend was observed in the 25% conditions (b, f), although it did not reach statistical significance. Behavior very similar comparable between groups for 50% trials (c, g). On 75% trials (d, h), High Proneness participants were significantly less likely to report hearing the voice, but this difference was not evident between AH− and AH+ groups. There was a significant block-by-group interaction for both groupings. *P < .05; ***P < .001. Note: CH, conditioned hallucination.

We also performed additional analyses to explore how detection rates at the 75% condition changed throughout the experiment (figures 3d and 3h). Repeated measures ANOVAs (supplementary tables S7 and S8) indicated a significant effect of the interaction between hallucination-proneness group and trial for both recent hallucinations (F_1,3166 = 23.8, P = 1.15 × 10⁻⁶) and general proneness (F_1,86 = 10.5, P = 1.22 × 10⁻³) measures.

These results suggest that while both recent and general hallucination-proneness correlate with CH rates, recent hallucination status most strongly relates to CH rate and general proneness has a stronger relationship with how the participant learns over time. To further disambiguate these effects, we fit behavioral data to a computational model describing latent states driving task performance.

High Prior Precision and Decision Noise Are Related to Hallucination Severity

We fit the parameters of a 3-tiered HGF model using behavior on the CH task (figure 4a). We found that higher CH rates correlated with greater weighting of prior expectations relative to incoming sensory evidence (ν; t₈₆ = 9.26, P = 1.47 × 10⁻¹⁵, R = 0.702). Given the differences in confidence between groups, we performed additional analyses demonstrating that ν also negatively correlated with confidence when reporting “no” on no-voice (F_1,86 = 8.24, P = 5.16 × 10⁻³, R = −0.30), 25% (F_1,86 = 6.75, P = .01, R = −0.27) and 50% trials (F_1,85 = 4.55, P = .036, R = −0.23) condition and when saying “yes” on 75% condition (F_1,86 = 5.64, P = .02, R = −0.25).

As part of our primary analysis, we show that those in the AH+ group relied more on priors (H₁ = 4.91, P = .03; figure 4c) and as a secondary analysis we show that reliance on priors increased as severity of recent hallucinations increased (t₈₆ = 2.89, P = 4.83 × 10⁻³, R = 0.280; supplementary figure S2a). However, we did not find a relationship between prior overweighting and general hallucination-proneness (Low- and High-Propensity: H₁ = 2.13, P = .145; LSHS total score: t₈₆ = 0.78, P = .439; figure 4f and supplementary figure S2c). In an exploratory analysis, we also found that those who had higher hallucination-proneness had significantly higher decision noise (β⁻¹; Low- and High-Propensity: H₁ = 8.61, P = 3.34 × 10⁻³; figure 4g) but not seen when comparing groups based on recent hallucinations (AH− and AH+: H₁ = 2.61, P = 8.44 × 10^–3; figure 4d). However, decision temperature increased with both general (LSHS total score: F_1,86 = 8.57, P = 4.38 × 10⁻³, R = 0.28; supplementary figure S2d) and recent hallucination severity scores (F_1,86 = 7.27, P = 8.44 × 10⁻³, R = 0.31; supplementary figure S2b). In general, belief-state trajectories did not differ between groups (figures 4b and 4e); however, we did find a significant effect of recent hallucination group on trial-wise beliefs (figure 4b, bottom, F_1,86 = 4.26, P = .04).

Sensory Threshold and Prior Weighting Differentially Predict Hallucination Severity and Hallucination-Proneness

We conducted an exploratory multiple regression analysis to determine if sensory threshold and prior weighting (ν) were independent predictors of recent hallucination severity and hallucination-proneness. As can be seen in supplementary tables S9 and S10, both threshold and (ν) independently predicted recent hallucination severity (total AHRS score) with a model r² of 0.19. Threshold, but not ν, predicted hallucination-proneness (total LSHS score), with a model r² of 0.12. This finding suggests that both reduced sensory precision and increased prior weighting are independently related to hallucination severity, consistent with our prior modeling work,^1,2 while decreased sensitivity is related to lifetime proneness, rather than hallucination state, which we have shown to be related to prior precision.¹³

Exploratory Analyses of Metacognitive Sensitivity and Efficiency

Given the general reduction of confidence observed among the hallucination-prone group, we performed exploratory analyses of metacognitive sensitivity as ascertained by meta-dʹ.^46,47 In brief, meta-dʹ is a bias-free measure of metacognition as derived from a Signal Detection Theory (SDT) framework (ie, the amount of signal available for metacognitive processing secondary to the first-order task of stimulus classification). Meta-dʹ is an analogue to the classical sensitivity index known as dʹ (meta-dʹ reflects a participant’s ability to discriminate between correct and incorrect responses in terms of subjective confidence while also controlling for overall task performance as determined by dʹ. In addition to meta-dʹ, because both dʹ and meta-dʹ are calculated in the same signal-to-noise ratio units of measurement, the 2 measures can be directly compared. We computed meta-dʹ, an M-ratio (meta-dʹ/dʹ), a measure of metacognitive efficiency, and an M-difference score (meta-dʹ − dʹ) for each participant at each condition level (overall, 75%, 50%, and 25%) to compare metacognitive efficiency between groups. We did not calculate these metrics at the 0% condition, as mathematically, this is the negation of overall meta-dʹ (ie, meta-dʹ of all no-signal trials as compared with all signal trials). The M-ratio can be conceptualized as the amount of sensory information (as determined by dʹ) ie utilized when making metacognitive judgements, with “optimal” metacognitive sensitivity being an M-ratio of 1. An M-ratio greater than 1 suggests that awareness is more precise than accuracy, and this may be reflective of subjects’ engagement of post-decisional processes, use of information outside of the task paradigm (eg, heuristics), or when decisions are made under time pressure.^48,49 Relatedly, the M-difference score is distance in terms of maximum utility of sensory information of meta-dʹ and dʹ, with “optimal” M-difference being 0.

In our modeling of meta-dʹ, we found no significant group differences in metacognitive sensitivity or efficiency at any condition between the AH+ and AH− groups (supplementary table S11). However, when comparing the Low and High hallucination-prone groups, we found significant group differences in overall meta-dʹ between the Low- and High-Propensity groups with the High group showing increased metacognitive sensitivity as compared with the Low-Propensity group. Moreover, we found that compared with individuals in the Low-Propensity group, individuals with high hallucination-proneness exhibited increased metacognitive efficiency as reflected by M-ratios, both overall and at the 75% and 50% conditions. We did not observe any group differences in M-difference scores at any condition between the Low- and High-Propensity groups.

Discussion

The primary purpose of this study was to determine if hallucination-proneness is related to a tendency to overweight one’s perceptual priors as measured by the CH task in the general population. We found increased CH rates in participants with increased hallucination-proneness (as measured by the LSHS) and more severe recent hallucinations (as measured by the AHRS), consistent with past work demonstrating a tendency of individuals with hallucinations to over-rely on priors in clinical and targeted nonclinical samples^{5,12,13,50,51} and with continuum models of psychosis.^52,53

Some of these results may provide interesting mechanistic clues about the development of hallucinations. As we and others have argued, a relative overweighting of priors may be the result of primary or secondary processes, the latter being viewed as a compensatory response to unreliable or noisy incoming sensory evidence.⁸ In this general population sample, we again demonstrated a link between higher relative prior weighting (ν) and increased recent hallucination severity scores at the time of assessment. Because this term is defined relative to the weighting afforded to sensory evidence, ν alone cannot tell us about the absolute weighting of priors or sensory evidence in isolation. However, other results provide some clues as to the status of both priors and sensory precision.

Participants with higher hallucination-proneness have an increased QUEST-derived threshold for initial detection of the most audible stimulus, suggesting lower sensitivity to sensory evidence. Second, those with higher hallucination-proneness demonstrate reduced detection rates for more easily audible stimuli. Beyond simple detection, participants with higher hallucination-proneness have reduced confidence in their responses about stimulus detection. These results are consistent with a well-established literature demonstrating sensory deficits in schizophrenia (see ⁵⁴ for review), and suggest reduced sensory precision—a reduced capacity to effectively use stimuli present in the outside world during perceptual decision-making. This reduced capacity could be due to reduced precision engendered by neurobiological changes such as white matter deterioration, which can be present in people at risk for psychosis.^55,56

Some results also hint at increased absolute prior precision in voice-hearers: high-proneness and AH+ groups begin and end the experiment with higher CH rates than low-proneness and AH− groups (figure 2), providing evidence for both higher prior precision at the start of the experiment and slower belief updating with new evidence. These differences are also reflected in slower belief updating on HGF analysis (figure 4), although primarily for AH+/AH− groups. These groups also report proportionally higher rates of detection when stimuli become more difficult to detect, indicating that their perceptual decision-making is more readily driven by expectations than their nonhallucinating counterparts.

It is possible that a general tendency toward over-reporting of detection may be the driver of increased CH rates in the higher-proneness groups. However, this appears unlikely: the higher-proneness group was calculated to have a higher threshold using QUEST; over-reporting of detection would have led to a lower calculated threshold, as the adaptive procedure decreases stimulus intensity with each detection reported. It is also possible that this higher threshold estimate could drive a general tendency toward over-reporting detection in the main experiment. However, we do not see a pattern of overall over-endorsement across trial types in the high-proneness group; rather they have reduced detection rates on the task with more audible stimuli.

Consistent with previous results,¹³ we have demonstrated that CH and prior overweighting are related to a greater extent to recent severity of hallucinations than to general propensity. This finding could be due to the state-sensitive nature of the ν parameter¹³; given that this was a general population sample, prior overweighting would be expected to be less significant, overall, than in a sample with psychosis or supra-threshold hallucinatory symptoms. As such, it might be reasonable for it to only become significantly elevated in the context of recent hallucinatory experiences, rather than general proneness.

Some of the findings, such as overall reduced confidence, increased detection thresholds, and noisier decision-making in the high-proneness groups, differ from results derived from previously published, tone-based versions of the CH task.^5,13 Without a direct comparison between the 2, it is difficult to say whether these differences in performance are due to the nature of the stimuli being used, but such a difference could point to interesting new directions for research: it may be that different subgroups of psychosis proneness (and psychosis) exist, in which different levels of the sensory processing hierarchy are affected. Alternatively, it may also be possible that different deficits are dominant at different stages of illness, with hallucination-prone groups presumably most closely resembling those in the earliest illness stages. In future planned work, we will compare these 2 tasks (eg, linguistic and tone) in the same groups to better assess these differences; we also discuss the importance of longitudinal studies below. Regarding the observation of higher decision noise in hallucination-prone groups, this may be consistent with recent findings of a link between decision noise and diminished cognitive capacity among nonclinical voice-hearers.⁵⁷

It is interesting to note that in previous work with clinical samples, as well as those with threshold-level symptoms but no disease, we observed an increased confidence on CH trials.^5,13 In this work, we relatedly observe reduced confidence on correct-rejection trials. This is consistent with past findings: more highly precise priors would cause one’s posterior (or final percept) to be closer to that produced in veridical stimulus-present trials, resulting in lower confidence in correctly reporting the absence of a target. We should not over-interpret this finding; however, it may be that this divergence in confidence in reporting false alarms and correct rejections reflects a divergence in reality testing between clinical and nonclinical voice-hearers.

Interestingly, despite observed lower confidence, participants with higher hallucination-proneness exhibit increased overall metacognitive sensitivity and metacognitive efficiency, as measured by meta-dʹ and the M-ratio of meta-dʹ/dʹ (see supplementary table S11)^58,59. This suggests that while confidence was lower among participants with high hallucination-proneness, they were able to better discriminate between their correct and incorrect answers in terms of their subjective confidence (ie, participants were more confident in their correct answers and less confident in their incorrect answers). While this finding may seem counterintuitive, when examining the M-ratio at each stimulus threshold, we found that at the 75% condition (ie, the condition with the most robust sensory evidence available), participants with high hallucination-proneness exhibited an M-ratio that exceeded a value of 1, or the “optimal” value of metacognitive efficiency,⁶⁰ suggesting that individuals with high hallucination-proneness have increased metacognitive awareness (meta-dʹ) as compared with their accuracy on the task (dʹ).⁶¹ Prior work has theorized that individuals who exhibit greater metacognitive awareness compared with task accuracy may be utilizing information other than the available sensory evidence, possibly including heuristics,⁶² or engaging in post-decisional processes that may affect metacognition.⁶³ These findings may suggest that rather than an “enhanced” use of sensory information to inform metacognitive awareness during the CH task, utilization of sensory evidence may be altered or supplemented by extraneous information or decisional processes outside of the confines of the task ie not apparent in individuals with low hallucination-proneness.

While we have highlighted the role of hyper-precise priors in the generation of hallucinations, it is important to note that other work has focused on an alternative explanation: that relatively imprecise prior beliefs underlie psychotic symptoms. Fletcher and Frith⁶⁴ suggest that overly precise prediction errors ascending the perceptual hierarchy could force changes at higher levels to accommodate or explain away the prediction errors, resulting in false beliefs. In this conceptualization, it is the relative weakness of prior beliefs in the face of overly precise ascending prediction errors, rather than overly precise priors, which drive aberrant belief formation, most obviously relevant to formation of delusions. Others^46,47 have offered a nuanced version of this perspective, with the suggestion that accounts of hypo- or hyper-precise priors might be reconciled by understanding that different priors may underlie different elements of psychotic experience (eg, delusions and hallucinations) and that the computational underpinnings of psychotic experiences might be better understood in the context of dysfunctional interactions between different layers within a hierarchical information processing system. Our work has emphasized the potential importance of disease stage, with different stages having potentially different computational characterizations.⁸ Haarsma et al aptly note the utility of laminar fMRI to more directly test these ideas and disentangle top-down and bottom-up factors leading to psychotic symptoms.⁴⁸

There are some limitations to the present work. There is a tradeoff between the convenience of a Mechanical-Turk-derived sample and the limitations of accurate self-report of clinical and symptom history.^49,65,66 This is potentially relevant, as there were a number of participants who had low AHRS and LSHS scores and yet had high CH rates, which were outliers compared with the other participants in the low CH group. It is possible that these participants are simply prone to high CH rates via a different mechanism than hallucination-proneness, that they did not respond to the test accurately in a manner our quality control does not measure, or that they actually were in the high-proneness groups but did not provide accurate self-report data. In addition, this experiment did not record any neural correlates of the behavior observed, limiting our ability to replicate hypothesized links to neurobiology. Additionally, because we define CH as a report of detection in the absence of a stimulus, it is possible that a conditioned motor response may be contributing to the group differences observed. We are confident from past work on the engagement of sensory cortices during CH that the behavior cannot be fully attributed to conditioned motor response.⁵ Additionally, in the current work, responses differ by auditory stimulus intensity presentation while the visual stimulus remains invariant, suggesting a perceptual judgment is being made.

It is possible that a greater propensity to hear voices in noise, unrelated to the conditioning in the experiment, may be helping to drive differences in CH rate between groups. As demonstrated in figure 3, the cumulative CH rate decreases as time (and the number of no-voice trials) increases. This suggests that a prior related to the association of auditory target with the visual cue is established and then weakens over time as contrary evidence is presented in the form of no-voice trials. We would not expect to see this effect if CH were driven by a general tendency to hear voices in noise. However, this possibility could be further examined in future work by adding trials containing auditory targets not previously associated with the visual cue.

There are several future directions that could be considered as a result of this work. One limitation of the HGF as instantiated here is that it does not directly model participant confidence in their sensory reports. Confidence reports have been useful in characterizing hallucinatory phenomena in nonhuman animals,⁶⁷ and including confidence in our modeling may help with translation across paradigms and species. Future work could be undertaken to correlate these behavioral results with neural measures. For example, proposed correlates of reduced sensory precision could be related to measures of white and gray matter integrity as well as functional connectivity using structural and functional imaging. While the nonsense words used in this experiment are closer to spoken language than the tones used in our previous work, they are still not fully ecologically valid stimuli. That said, the use of complex linguistic stimuli may reveal important differences even without semantic meaning attached: if, eg, some voice-hearers are susceptible to linguistic CHs and not to tone-based CHs, this may suggest the presence of subgroups within the voice-hearing population with information-processing disruptions at different levels of the auditory processing hierarchy. We are currently analyzing a related dataset to determine whether this is the case. In future work, we plan to use real spoken words as stimuli to probe the effect of semantics and affective valence on CH; however, the nonsense words tested here will serve as an important control condition in future work. Another future direction for testing would be to modify the intensity of white noise used. Although we manipulate signal-to-noise ratio here by altering the intensity of the target stimulus embedded in white noise, manipulation of the noise itself could help to isolate the role of auditory cortical excitability and incoming sensory noise in hallucinations.

Taken together, our results demonstrate that hallucination-proneness and hallucination severity are related to susceptibility to CH and attendant prior hyper-precision in the general population. Because of the sensitivity of the CH task in this population, similar tools may be developed to screen for or investigate factors driving subclinical psychosis-like phenomena in broad, diverse samples. We also demonstrate for the first time that hallucination-prone individuals may exhibit sensory disturbances that are potentially causally related to their tendency to over-trust their priors. If these results are confirmed and extended using novel paradigms, they may help inform causal models of the development of psychotic symptoms.

Funding

This work was conducted at the Connecticut Mental Health Center and was funded in part by the State of Connecticut, Department of Mental Health and Addiction Services, but this publication does not express the views of the Department of Mental Health and Addiction Services or the State of Connecticut. The views and opinions expressed are those of the authors. ARP is supported by a K23 Career Development Award, R21, and two R01s from the National Institute of Mental Health (K23MH115252-01A1; 5R21MH122940-02; R01MH129721; R01MH131768), by a Career Award for Medical Scientists from the Burroughs-Wellcome Fund, a Carol and Eugene Ludwig Award for Early Career Research, and by the Yale Department of Psychiatry and the Yale School of Medicine. EAF is supported by the Yale Doctoral Fellowship in Clinical and Community Psychology, through the Yale Department of Psychiatry and Yale School of Medicine.