Cross-modal cue integration is adaptively advantageous given the likelihood that real-world events perturb multiple sensory systems; and is sensible from a computational standpoint, as uncertainty in one modality can be reduced by combining with signals from others, increasing the overall precision of a sensory percept. However, perceptual binding is not straightforward because sensory signals vary continuously, instantaneously and unpredictably. Related inferences are therefore necessarily probabilistic. As inter-sensory inconsistencies are likely abundant, perceptual decisions should take into account stimulus-specific uncertainty. Indeed, humans have been consistently shown to fuse multisensory data in a manner that reflects the relative reliability of the information received in line with canonical ideal-observer models including Bayesian decision theory and maximum-likelihood models (Green and Swets, 1966; Ernst and Banks, 2002; van Beers et al., 2002; Alais and Burr, 2004).

The Bayesian perspective provides a general inferential schema to understand brain function. With respect to perception, our experiences can be understood as the most likely explanation of the world (determined via posterior probability distributions), with these inferences biased by our own expectations (or priors reflecting past inferences) as well as data from available stimuli (the conditional likelihood of the data given a model). Critically, these computations lead to updated posterior probability distributions about the world which then serve to adjust future expectations. The fundamental units are prediction error (on expected sensory input) and expected value (of actions executed in the environment), that can be modeled as free energy in a general optimization theory of perception and action (Feldman and Friston, 2010; Friston, 2010).

Speech perception presents us with two illusory phenomena whose occurrence illustrates the role of Bayesian inferences in multisensory fusion. Firstly, the ventriloquist illusion (that an inanimate puppet is speaking) occurs when a ventriloquist produces speech sounds while simultaneously moving a puppet's lips instead of his/her own. This robust effect denotes inaccurate source localization on account of the temporal proximity of the ventriloquist's speech and the puppet's lip movements. The illusory percept can be understood as the consequence of the expectation that multisensory stimuli coincident within specific spatio-temporal limits emanate from a common source, as is supported by the finding that the strength of the ventriloquism effect decreases as spatial and temporal disparity increase (Wallace et al., 2004). The observation that the ventriloquism percept can be increased when stimuli are more spatio-temporally coincident (Radeau, 1994) implies dynamic updating of top-down priors as previous conceptions are revised. Furthermore, the elevated weighting of visual localization cues can be explained by the modality-appropriateness hypothesis that weighting adjustments reflect our resolving capabilities within each sensory modality (Welch and Warren, 1980). Thus, the illusory interpretation is judged the most likely (and therefore optimal) explanation but is erroneous for the idiosyncratic and ambiguous conditions manipulated by the ventriloquist. This notion of illusory experience as a Bayes-optimal inference is elegantly formalized elsewhere (Brown and Friston, 2012).

Another illusory (yet optimal) inference during speech perception was first reported by McGurk and MacDonald (1976), who found that when viewing edited movie clips of an actor articulating one syllable in synchronization with the soundtrack of other (specific) syllables, individuals often perceive a syllable incongruent with either the visual or auditory input (McGurk and MacDonald, 1976). For example, a visual /ga/ with an auditory /ba/ is perceived as /da/; and a visual /ka/ with an auditory /pa/ elicits the perception of /ta/. This perceptual effect can be understood, similarly to the ventriloquist illusion, as reflecting the most plausible auditory percept under the assumption of a common signal source (a single articulating agent), when spatio-temporal multi-modal constraints are met.

These examples suggest that priors are estimated in multiple, hierarchically-arranged contexts: there is a general prior that lip movements and sounds usually come from the same person, which has the effect of inducing integration of visual and auditory information into a single percept; furthermore, priors pertaining to information presented in one modality (e.g., vision) have the potential to influence interpretation of data presented in another modality (e.g., audition); and particularly relevantly to the McGurk effect, further priors dictate the extent to which particular multimodal confluences shape perception. These putative priors accordingly provide explanation for illusions being systematically associated with certain audio-visual combinations but not others. If visual and auditory data are presented simultaneously and are therefore more likely to emanate from a common source (than contradictory independent sources), and their concurrent posterior probability distributions overlap with those of another explanation of the data, then favoring that alternative explanation is probabilistically advantageous. By contrast, if an aspect of the auditory or visual data is more distinct, its posterior distribution will be more precisely defined (i.e., there will be reduced variance around its maximum likelihood estimate), and that cue will have increased weight in the perceptual decision, which may in turn over-ride the prior that the inputs are commonly derived.

Several core features of schizophrenia can be understood in terms of defective Bayesian processes (Corlett et al., 2009; Fletcher and Frith, 2009). Abnormal interactions between prior beliefs and presented sensations plausibly hinder attempts to prioritize, learn about and understand the world (Gray et al., 1991); more dramatically, over-powering priors may cause the experience of a non-existent stimulus (hallucination), or abnormal bottom-up signaling may instill a conventionally-irrelevant event with enhanced significance (laying the foundation for a delusional belief) (Fletcher and Frith, 2009; Joyce et al., 2013). Pervasive deficits in information processing—including those relating to multisensory integration—further intimate impaired Bayesian inferences in the disorder (Heinz and Schlagenhauf, 2010; Shergill et al., 2014). Most relevantly here are previous observations that individuals with schizophrenia are less likely to perceive the McGurk effect (de Gelder et al., 2005; Pearl et al., 2009), although it is noteworthy that preserved multisensory fusion also been reported (Myslobodsky et al., 1992; Martin et al., 2013).

Bayesian models of the world require two fundamental inferences: updating a representation of the state of the world; and a judgment about the associated level of uncertainty (Friston et al., 2012). Dopamine has been proposed to play a role in the top-down specification of priors (Scott et al., 2007; Corlett et al., 2009), coding uncertainty by means of synaptic gain control (Friston et al., 2012). Gain—the rate at which neuronal output changes in relation to the strength of driving inputs—fundamentally promotes the amplification of contextually-relevant signals and the consequent reduction of noise in the inputs, which in turn imparts advantages for optimizing signal detection. This can play diverse computational roles; but in terms of multisensory integration is thought to promote appropriate binding by synchronizing event-related oscillatory activity through the coordinated action of fast-spiking inhibitory interneurons (Phillips and Silverstein, 2013).

Kapur's (2003) notion that delusions are formed when spontaneous, phasic striatal dopamine firing imbues salience to contextually-irrelevant stimuli presents a mechanistic link between Bayesian cognition and schizophrenic symptomatology. Individuals with schizophrenia exhibit an elevated baseline dopaminergic state (Laruelle et al., 1996), but reduced event-related, context-relevant dopamine firing (Kapur, 2003; Heinz and Schlagenhauf, 2010). The latter supports the notion that abnormal gain control and prior specification explain the pervasive sensory and decision-making deficits in the disorder (Phillips and Silverstein, 2013). This pathological firing can explain reduced susceptibility to illusory multisensory stimuli in schizophrenia since inconsistent dopamine-coded, event-specific precision (Friston, 2005) potentially reduces the likelihood of appropriately binding encountered stimuli from discrete modalities. Relatedly, it has been shown that dopamine agonists facilitate visual discriminative performance and enhance perceived sensory acuity; and that this improvement is associated with greater subjective feelings of event-related confidence (Fillmore et al., 2005; Lou et al., 2011). The investigation of the relationship between multisensory integration and metacognitive confidence judgments in schizophrenia provides an accessible means of increasing our understanding of the associated role of dopaminergic dysregulation. However, investigating dopaminergic modulation in association with pathology is associated with numerous confounds. Pharmacological manipulation of dopaminergic processing in healthy individuals provides an alternative investigative model.

Central to accounts of perception as probabilistic inference is their dynamic updating, which occurs on multiple timescales. In light of modality-specific developmental and degenerative changes in sensory acuity through the lifespan (Gordon-Salant, 2005), the modality-appropriateness hypothesis, and the role of cumulative experience, age is a likely moderator of the efficiency of multisensory integration. McGurk perception has been previously reported to vary with age, with older adults more likely to experience the McGurk effect than younger adults (Setti et al., 2013). Given characteristic developmental disruption in schizophrenia (Murray and Lewis, 1987), age may also differentially affect multisensory integration in individuals with schizophrenia. This is intimated by the findings that significant differences in illusory McGurk perception have been previously observed in adolescents but not older individuals (Pearl et al., 2009).

The primary aim of this work was to quantify audio-visual performance and related metacognitive judgments to test four explicit hypotheses. First, lower rates of illusory perception were predicted in individuals with schizophrenia as compared with healthy individuals. This prediction was made on account of the converging lines of evidence suggesting perceptual aberrance, disturbances to probabilistic inference, and previous abnormalities in multisensory fusion in schizophrenia (Heinz and Schlagenhauf, 2010; Phillips and Silverstein, 2013), and the notion that illusory perception represents the optimal probabilistic inference. Second, and complementarily, on account of a putative reduction in the acuity of multisensory integration in association with reduced dopaminergic activity, it was hypothesized that illusory McGurk perception would be reduced in healthy individuals following pharmacologically-induced dopamine blockade. In light of the idea that probabilistic inferences are dynamically updated by experience, it was predicted that schizophrenia-related abnormalities would increase in line with illness chronicity. Finally, the influence of illness, drug and age on metacognitive evaluation of audio-visual performance during the McGurk task was assessed to test the fourth hypothesis that these factors shape higher-order subjective experiences associated with multisensory fusion in addition to the perceptual experiences themselves.

Principal findings of this behavioral study of audio-visual integration were: that individuals with schizophrenia were less likely to perceive the McGurk illusion than healthy controls; and that, in these individuals but not control subjects, age significantly and inversely predicted the rate of illusory experience. The latter finding, in limiting associations with age to the patient group, implies a role for illness chronicity. This is substantiated by the observation that illness duration robustly predicted rates of illusory perception. Overall, these findings are consistent with the proposal of a Bayesian-like probabilistic inference process being at work in the McGurk effect, modification of this process by experience, and disturbance to this system in schizophrenia.

Reduced susceptibility to the McGurk effect has been inconsistently reported in individuals with schizophrenia (Myslobodsky et al., 1992; Surguladze et al., 2001; Pearl et al., 2009; Szycik et al., 2009; Martin et al., 2013). Several studies reporting null findings were underpowered to detect effects of small magnitude on account of their small samples. The inconsistency in the literature feasibly reflects these small-to-moderate sample sizes, heterogeneity in the clinical phenotype of the disorder and between-study variation in the sensory stimuli presented. However, as considerable inter-subject variability exists in susceptibility to illusions (as alluded to by previous analyses investigating perception in subsets of responders and non-responders), definitively establishing whether susceptibility is affected in schizophrenia fundamentally depends on identifying and accounting for factors controlling this variability.

The current observation that individuals with schizophrenia were less likely to bind auditory and visual information may seem at first glance incompatible with the considerable literature that individuals with the disorder require greater temporal asynchrony between stimuli to dissociate them (Foucher et al., 2007; Giersch et al., 2009; Schmidt et al., 2011). However, the McGurk effect relies on implicit multisensory binding, while judgments of temporal asynchrony are precisely explicit. Differences in the preservation of these capabilities are expected in light of the independence of implicit and explicit judgments and previously-observed differences in their impairment in schizophrenia (Lalanne et al., 2012a,b). Furthermore, irrespective of the conceptual independence of these judgment types, reduced susceptibility to illusions and reduced performance in temporal asynchrony tasks can be reasonably interpreted as deviation from the behavior of an optimal observer.

Disturbances to multisensory perception have also been observed in relation to emotional processing in schizophrenia. de Gelder et al. (2003, 2005) found in emotion recognition tasks that individuals with schizophrenia were less likely than healthy controls to benefit from confirmatory vocal evidence when assessing facial expression but also exhibited exaggerated face effects when judging vocal emotion. This denotes non-conformity with the general principle that responses to ambiguous stimuli are more influenced by a second stimulus even when this stimulus is extraneous (Massaro, 1998). Interestingly, the same individuals showed inconsistent voice categorization in even the least challenging task conditions suggesting high-level auditory uncertainty. It is therefore a vital future objective to dissociate deficiencies in perceptual decision making from those directly related to multisensory integration.

In spite of the previous observation that varying attentional load does not impact proficiency of multisensory fusion (Vroomen et al., 2001), it is feasible that the observed between-group differences have in part an attentional foundation. Individuals with schizophrenia have been replicably shown to exhibit reduced visual focus on socially-important face regions such as the mouth and eyes (Loughland et al., 2002a,b); which could underlie their performance abnormalities in this task, in addition to other social deficits. Future work would therefore be greatly strengthened by the collection of eye-tracking measurements during task performance. In fact, significant associations between poor sensory integration and eye-tracking disorders have been previously reported in schizophrenia (Ross et al., 1998).

Age interestingly moderated audio-visual performance, predicting the rate at which individuals perceived the McGurk effect. However, while significant associations with age were noted as a whole-sample characteristic, they were evidently stronger in the schizophrenia group, which suggests that the age-related decrements observed in the schizophrenia group are attributable to long-term illness rather than the ageing process itself. This notion is strengthened by the highly-significant relationship between illness chronicity and the rate of illusory perception seen in these individuals. An alternative perspective is that reduced illusory perception in individuals that have been longer diagnosed with schizophrenia is attributable to progressive abnormalities in brain structure and function (van Haren et al., 2008), which in turn impair the systems on which multimodal integration is dependent. The cumulative role of other interacting factors such as long-term medication is also unclear, although medication dosage was not related to perceptual performance in the current investigation.

The current findings partially contradict those of Pearl et al. (2009), who, despite similarly reporting diminished susceptibility to the McGurk effect in individuals with schizophrenia, observed an alternative age effect: healthy adolescents experienced the illusion more frequently than healthy adults, but perception did not differ by age in their stratified schizophrenia group. Pearl et al.'s (2009) findings are in line with the modality-appropriateness hypothesis, which suggests that the weighting of sensory signals differs through life, presumably on account of sensory-neural degeneration in later adulthood (Gordon-Salant, 2005). The McGurk effect is seemingly contingent on an individual possessing requisite acuity of both visual and auditory processing. Inter-subject variation in sight and/or hearing capacity may underlie the differing susceptibility to illusory perception exhibited between healthy individuals; although this remains a speculative interpretation in the absence of explicit sense testing. The discrepancy between our healthy-sample findings and those of Pearl et al. (2009) may be partially attributable to the previous study recruiting individuals whose ages ranged beyond that of those included in the current sample, and who were therefore more likely to exhibit sensory decline. Furthermore, the mean illness duration of the current schizophrenia group (13.9 ± 8.7 years) was greater than that of the previous study (7.2 ± 6.2 years), which potentially enhanced our sensitivity to detect effects of illness chronicity.

Another, alternative relationship between age and susceptibility to the McGurk effect was observed by Setti et al. (2013), who found increased illusory perception in a sample of older adults (mean age: 65 years) compared with younger adults (mean age: 22 years) and interpreted this increase in terms of perceptual decline in the older group. The current findings are compatible with these observations, on account of the differences in age distribution between the study samples. However, the interpretation that McGurk illusory perception is more likely in individuals with deficient sensory performance is somewhat at odds with the view that it represents optimal function. Cienkowski and Carney (2002) also reported evidence of normal-to-enhanced McGurk effects in a small sample of older as compared with younger adults. This study elegantly measured their participants auditory (but not visual) capabilities but its conclusions are somewhat limited by a small sample size. The effects of age-related sensory decline on illusory perception remain to be definitively characterized.

Deficits in the formation and updating of expectations potentially underlie schizophrenia-related abnormalities in reward processing (Heinz and Schlagenhauf, 2010; White et al., 2013); sensorimotor predictive coding (Shergill et al., 2005, 2014); and wide-ranging features of perception (Phillips and Silverstein, 2013), including multisensory perception. The impact of expectations on all of these processes highlights their fundamental importance for adaptive behavior; and intimates pervasive difficulties for individuals in whom this system breaks down. Nevertheless, in light of the heterogeneity in the clinical presentation of schizophrenia, the dynamic nature of probabilistic inference and the numerous domains in which it has been found to be abnormal in schizophrenia, it remains a considerable challenge to isolate any specific putative disturbance in the disorder. Computational modeling offers the most viable current means of characterizing the relevant functional components. Investigation of specific inference-related parameters (for example, the temporal trade-off between existing and new evidence) offers a window on the clinical features of schizophrenia (Joyce et al., 2013). Similarly, such models may permit future dissociation of aberrances in unimodal perception and those of defective integration. However, this may be a false distinction in Bayesian terms (in some circumstances), since it is possible that a single pathophysiological abnormality could echo through the inherently hierarchical system impairing both processes.

A reduced McGurk_illusory rate, indexing an impairment to multisensory predictive inference, was expected in individuals with schizophrenia with prominent positive symptoms, given Bayesian accounts of these symptoms (Fletcher and Frith, 2009). However, while a trend toward this putative relationship was indeed observed, more robust was the positive association between negative symptoms and the McGurk_illusory rate. Impaired probabilistic inference can be reasonably hypothesized to underlie negative symptoms since continually encountering seemingly-unreliable sensory data could plausibly hinder decisions, render actions inconsequential and ultimately induce the paucities of thought and movement associated with negative symptoms. Paradoxically, in the current dataset individuals with prominent negative symptoms were more susceptible to the illusion. Thus, while the Bayesian account offers a compelling procedural explanation for multisensory integration (Seilheimer et al., 2014) and its abnormality in schizophrenia, alternative explanations should also be considered. For instance, dysconnectivity of brain structure, function, and mental experience have been diversely observed in schizophrenia (Fitzsimmons et al., 2013; Palaniyappan et al., 2013; White et al., 2013), and may fundamentally impair the inter-regional synchronization of activity upon which effective multisensory integration depends. Alternatively, impaired multisensory fusion in schizophrenia could potentially be attributable to modality-specific attentional deficits in these individuals.

The investigation of metacognitive appraisal of multisensory perception yielded consistent effects of experimental condition, with lower confidence ratings in association with the ambiguous illusion-eliciting stimuli as compared with control stimuli and a tendency for this effect to be less pronounced in individuals with schizophrenia than healthy individuals. This result can be interpreted as the control conditions proving more challenging for the individuals with schizophrenia; and healthy individuals exhibiting enhanced awareness of the intrinsic ambiguity of cue information. Nevertheless, main group effects were not significant. The relationship between confidence and dopamine potentially holds the key for understanding several core clinical features of schizophrenia. Delusional beliefs are confidently held without logical substantiation; it is possible that belief formation governed by phasic dopamine release is aberrant in schizophrenia due to dysregulated firing in association with conventionally non-salient events (Kapur, 2003); and further that the hyper-dopaminergic tonic dopamine activation produces the immutable, high-confidence associated these beliefs. Failure to appropriately update expectations could explain both the liberal acceptance thresholds seen during decision making in schizophrenia (Moritz et al., 2007), and cognitive predispositions such as the bias against disconfirmatory evidence (Moritz and Woodward, 2006; Woodward et al., 2007).

We report for the most part null effects of amisulpride administration on multisensory performance. This may reflect in part the relative timing of dosing and task performance. The initial T_max of amisulpride is associated with lower plasma concentration than the second (Kudris et al., 2011); it is therefore possible that the current investigation was not advantageously placed to investigate dopaminergic modulation of perception.

A limitation of the current protocol is the differential use of placebo conditions between the schizophrenia and healthy groups. While the concordance of our results with previous investigations of the McGurk effect in schizophrenia (de Gelder et al., 2005; Pearl et al., 2009) suggest that this is not the sole driver of our between-group differences, previous links between placebo responses and dopamine warrant discussion. Striatal dopamine has been shown to be positively associated with both anticipatory responses during reward learning (with the latter indices of the effects of Bayesian priors), and the level of placebo responses to preparations believed to have anti-nociceptive properties (Scott et al., 2007). Direct investigation of placebo effects on multisensory integration is therefore a worthwhile line of future research. In addition, it would be useful for future work on multisensory binding in schizophrenia to use Bayesian models to evaluate the effects of shifting stimulus onset asynchronies between auditory and visual information to ascertain the weighting of reciprocal effects of each modality on the other. A further potential limitation of this work is the imperfect between-group IQ matching. However, the WASI indexes current rather than premorbid IQ and the current sample exhibited similar (if slightly higher) scores to individuals with schizophrenia tested previously with the same scale (Weickert et al., 2000; Barrett et al., 2009). Together, these findings demonstrate the test's sensitivity to the cognitive deficits characteristic of the schizophrenic clinical phenotype. Furthermore, the control and patient groups were well matched in terms of age, sex and socioeconomic background, which serves to validate the recruitment procedures used.

While the focus of the current work is on dopamine, whose effects appear to enforce top-down adaptation of signaling, there is evidence that other neuromodulatory transmitters play complementary roles in Bayes-guided processing. Acetylcholine, for instance, has been recently shown to boost bottom-up signaling in predictable environmental conditions (Moran et al., 2013). In light of the inherent interdependence of these and other neurotransmitter systems, and the hierarchical and reciprocal nature of probabilistic inference (Friston and Kiebel, 2009; Feldman and Friston, 2010; Friston et al., 2012), determination of transmitter-specific computational effects and related interactions will also greatly improve our understanding of related cerebral mechanisms.

In summary, this work further demonstrates the utility of illusion-provoking scenarios for the investigation of information-processing abnormalities in schizophrenia (Dima et al., 2009; White and Shergill, 2012). In addition to confirming attenuated rates of illusory experience in the disorder, which are interpreted as an indication of sub-optimal inference, we present data on the subjective milieu surrounding perception of these ambiguous sensory situations. The incorporation of metacognitive appraisals provides a window on the dynamic effects of expectation not only on perception per se but also inferences drawn from and about it. While alternative explanations cannot be unequivocally discounted, our findings add to the flourishing literature suggesting that schizophrenic pathophysiology is usefully framed in Bayesian terms.