Confidence in visual detection, familiarity and recollection judgements is preserved in schizophrenia spectrum disorder

Introduction

Confidence abnormalities in the form of overconfidence in errors in schizophrenia spectrum disorder have been documented in multiple cognitive domains, including memory, perception, and emotion recognition². Yet, the hierarchical level at which these abnormalities occur is still unclear. In line with the terminology proposed by Galvin and colleagues³, cognitive performance is referred to as first-order performance (i.e. how well one is able to detect or discriminate between probed stimuli), and metacognitive performance is referred to as second-order performance (i.e. how well one is able to discriminate between correct and incorrect responses). Properly quantifying metacognitive abilities requires controlling for variations of cognitive performance that are not metacognitive in nature^3–5. This concern is of particular relevance in schizophrenia, where cognitive deficits are well documented^6,7. In a meta-analysis we recently conducted¹, metaperception was mostly preserved when first-order performance was controlled for. Yet, conclusions about the metamemory deficit could not be drawn in this meta-analysis since the medium to large effect size resulted from studies where memory performance was not equated between patients and healthy controls (except for one study⁸). In these conditions, metamemory deficits were likely to be confounded with memory deficits.

To compare metaperceptual and metamemory deficits in individuals with schizophrenia while controlling for perceptual and memory deficits, we developed a novel experimental paradigm including three randomly interleaved perceptual and memory tasks attempting to experimentally match first-order performance at the intra-individual level across tasks, and to statistically control for performance at the inter-individual level.

We preregistered our main predictions based on current knowledge regarding the cognitive architecture of perception and memory and their impairments in schizophrenia (see⁹ for a meta-analysis). Individuals with schizophrenia typically have preserved performance in familiarity judgements (i.e. decontextualized memory¹⁰) but impaired performance in recollection judgments (i.e. episodic/recollection memory necessitating multimodal integration via hippocampal activity¹¹), which may be explained by impaired hippocampus recruitment¹² and hippocampal atrophy¹³. Our main preregistered hypothesis was that metamemory was globally more impaired than metaperception, assuming that previous reports of deficits in metamemory were not only driven by deficits taking place at the first-order level. Furthermore, since familiarity can be considered a perceptual-mnemonic process storing decontextualized perceptual elements¹⁴, we hypothesized domain-generality between perception and familiarity processes, and expected that meta-recollection would be specifically impaired. Besides this preregistered hypothesis, we explored the links between metacognitive performance and clinical traits such as positive, negative and disorganization syndromes.

Methods

The present design, hypotheses, and analyses were preregistered prior to data collection and analysis (https://osf.io/k4p79). Data and analysis scripts are available online (https://gitlab.com/nfaivre/metaface_scz_public).

Experimental design

A video description of each task is available online (https://gitlab.com/nfaivre/metaface_scz_public/-/tree/main/videos). All participants were naive to the purpose of the study, gave written informed consent in accordance with institutional guidelines and the Declaration of Helsinki, and received monetary compensation (10€ / h) except those participants under legal protection. The study was approved by the ethical committee Sud Méditerranée II on April the 3d 2020 (217 R01 MS1).

Stimuli

4000 copyright-free artificially generated faces were downloaded from the open platform https://generated.photos. Two independent observers screened the stimuli to exclude children’s faces, unrealistic faces, and faces with salient features (e.g. sunglasses, hats). The remaining 1700 male and 1700 female adult faces were converted to grayscale and equalized in contrast and luminosity (SHINE Matlab toolbox¹⁵). Each face was presented against a visual background noise consisting of its phase-scrambled version. The background was colorized in blue or red (balanced for luminosity) to provide a contextual cue. Size and gaze position were kept identical across all stimuli.

Procedure

Memory tasks

The familiarity and recollection tasks shared the same timeline (Figure 1). Each trial started with an encoding phase consisting of four successive face stimuli presented during 400ms each (random combination of 2 male and 2 female faces) on a blue or red background (context), with a 500 ms inter-stimulus interval. To avoid learning effects and familiarity confounds, each face was presented only once throughout the whole experiment. Following the encoding phase, the test phase consisted in presenting a fifth face on a gray background, and asking a task-specific question. In familiarity trials, the participant was asked to indicate whether the face had already been seen (80% of the trials, to obtain a uniform distribution across “stimulus strength” levels, see next paragraph) or not (20% of the trials); in recollection trials the fifth face was always a seen face (i.e. a face presented during the encoding phase), and the participant was asked whether the context of this stimulus was blue (80% of the trials) or not (20% of the trials) during the encoding phase. Participants provided their answers with a mouse click on “no” or “yes” buttons respectively displayed at the top left and top right of the screen.

• Download figure
• Open in new tab

Figure 1.

Experimental Design. Timeline of the familiarity, recollection and visual detection tasks. The timeline was identical in the familiarity and recollection task, except for the testing phase where the question was task-specific: “Already seen?” for familiarity, and “Blue context?” for recollection. No encoding took place in the visual detection task. In the present illustration, the correct answers to the familiarity, recollection and perceptual questions are respectively: “No”, “Yes”, and “No”. Lag is an ordinal variable corresponding to the temporal distance between the target stimulus and the test stimulus.

The difficulty of the familiarity and recollection tasks was manipulated by changing the serial position of the target stimulus during the encoding phase. Accordingly, there were four levels of stimulus strength - ranging from 1 to 4 -, corresponding to each of the four faces displayed sequentially within the encoding phase (Figure 1). Because this variable corresponds to the temporal distance between the target stimulus and the test stimulus, we refer to it as a “lag”. For instance, if the target face was the first face displayed during the encoding sequence, then the temporal distance between the target and the test was maximal, and the trial was categorized as “lag 4”. On the contrary, if the target was the last face of the encoding phase, the temporal distance between the target and the test was minimal, and the trial was categorized as “lag 1”. A fifth lag-level “lag 0” was used to indicate catch trials (20% of the trials): i.e. new faces in familiarity trials, faces presented in the red context in recollection trials.

Visual detection task

Participants had to indicate whether a face was present (80% of the trials) or not (catch trials: 20%). The face could be presented at four contrast levels, chosen to match performances obtained in the memory tasks for each of the four lags (See SI Figure S2 D). A fifth level - stimulus strength 0 - was used to tag catch trials: trials where no face was presented (20% of the trials). As for memory trials, participants provided their answer with a mouse click on “no” or “yes” buttons displayed at the top left and top right of the screen.

Trial exclusions

A time limit of 6 seconds was set on all trials to avoid differences in response rates between patients and controls. When the time limit was reached, an error-like sound was produced along with a visual warning in red characters asking participants to respond quicker. Proportions of non-responses were comparable between individuals with schizophrenia (mean ± SD: 2.91% ± 3.60%) and controls (mean ± SD: 2.32% ± 7.66%, BF = 0.26). These non-response trials were excluded from our analyses.

Confidence rating

For all three tasks, participants were asked to provide confidence judgments. After each first-order response (i.e. responses given to the familiarity, recollection and visual detection tasks) participants were asked to report their subjective confidence regarding the correctness of their decision by moving a slider with the mouse on a visual analog scale (see Figure 1) ranging from 0% (“Sure incorrect”) to 100% (“Sure correct”). The initial position of the cursor for each trial corresponded to 50% confidence (“Not sure”).

Structure of the experiment

This protocol aimed to match intra-individual performance across familiarity, recollection, and visual detection tasks. Participants were asked to perform two sessions of one hour each. Session 1 allowed us to measure memory performance at four difficulty levels (according to the variable “lag”, see Memory Tasks). We then matched perceptual performance to memory performance by determining four adequate contrast levels for the visual detection task for each participant (see SI for details). Thus, session 1 provided four levels of stimulus strength, i.e. 4 memory lags and 4 visual contrast levels, corresponding to matched performance for each participant. Based on these individual parameters, session 2 contained 10 blocks of 30 randomly interleaved trials (familiarity, recollection and visual detection task), totalizing 300 trials (100 trials per task), each followed by a confidence rating task. Task order and stimulus strength were randomized, so participants could not predict which task they were going to perform on each trial.

Importantly, this paradigm was designed to match first-order performance between tasks, which is convenient to compare metacognitive deficits across tasks. Although we also attempted to match first-order performance between groups, pilot experiments revealed this was not possible using adaptive staircases. Therefore, differences in task performance between groups were accounted for at the statistical level using the confidence efficiency metric¹⁶, taking advantage of our design with different levels of difficulty.

Results

First-order performance

Model 1a revealed that patients had lower performance than healthy controls in the visual detection, familiarity and recollection tasks, and these first-order deficits were similar across tasks (i.e. no first-order interactions, see Table 3, Figure 2A).

View this table:

• View inline
• View popup
• Download powerpoint

Table 3:

First-order deficits across tasks. We report posterior distributions’ summary statistics (mean and 95% Credible interval) along with Bayes factors.

• Download figure
• Open in new tab

Figure 2:

A. Hit rates (i.e., rates of “yes” responses following stimuli with stimulus strength > 0) across stimulus strengths in the visual detection (purple), familiarity (yellow), and recollection tasks (orange). Points and error bars indicate average accuracy and standard error of the mean, respectively; solid lines and shaded areas represent model fit mean and 95% confidence interval, respectively. B. False-alarm rates (i.e., rates of “yes” responses following stimuli with 0 stimulus strength) across groups. Points and error bars indicate average accuracy and standard error of the mean, respectively. Same color description as panel A.

Differences in performance were expected as task performance was not experimentally controlled between groups. However, our procedure was designed to match intra-individual performance across tasks. Accordingly, pairwise first-order task performances were similar among patients and among control participants (Table 3). This confirms that our procedure globally matched intra-individual performance across tasks, although it did not match intra-individual performance for each stimulus strength (see Table S1).

Patients and controls were sensitive to task manipulation of stimulus strength as indicated by a strong effect of stimulus strength in all tasks (See Table S1 and Figure 2A).

False alarms

Compared to healthy controls, patients had a similar false alarm rate in both the visual detection task (i.e., reporting seeing a face when none was presented: -0.22 [-1.02, 0.57], BF = 0.45), and in the familiarity task (i.e., reporting having seen the test face during the encoding phase when presented with a new face: 0.78 [-0.07, 1.64], BF = 2.21) but they committed significantly more false alarms in the recollection task (i.e., reporting having seen the test face in a given context during the encoding phase when presented in another context: 1.54 [0.61, 2.46], BF = 96.7) (Figure 2B).

Second-order performance

Confidence

Confidence levels were similar between patients and controls, except for the recollection task where patients were underconfident in correct responses (Table 1, confidence mean ± SD: 80.9 ± 10.9) compared to controls (confidence mean ± SD: 87.4 ± 10.0, t = -2.58, p < 0.05, BF = 4.04).

Metacognitive sensitivity

When quantifying metacognitive sensitivity as the slope between accuracy and confidence in mixed-effects logistic regressions (model 1a), individuals with schizophrenia were not found to underperform compared to healthy controls (Figure 3A). Although qualitatively, the results could suggest a metacognitive deficit in the visual detection task, the evidence was statistically inconclusive (−0.41 [-0.84, 0.01], BF = 1.33). By contrast, we obtained moderate evidence in favor of an absence of a deficit both in meta-familiarity (−0.24 [-0.59, 0.12], BF = 0.32), and meta-recollection (−0.13 [-0.51, 0.27], BF = 0.17). Moreover, there was no difference of deficit between tasks (Familiarity - Recollection: 0.11 [- 0.33, 0.56], BF = 0.18); Familiarity - Perception: 0.17 [- 0.26, 0.62], BF = 0.28; Recollection - Perception: 0.28 [- 0.18, 0.75], BF = 0.47). As discussed above, metacognitive sensitivity can be contaminated by differences in terms of first-order performance, which was only partially controlled in our paradigm. To estimate metacognitive performance independently of first-order performance, we turned to another metric called the confidence efficiency.

• Download figure
• Open in new tab

Figure 3:

A. Bayesian posterior distributions of differences of regression slope estimates between patients and controls (i.e. distributions of metacognitive deficits estimations): meta-perceptual difference (purple), meta-familiarity difference (yellow), meta-recollection difference (orange). Vertical dashed line (estimate = 0) represents no difference between patients and controls. Horizontal colored bars indicate 95% credible intervals. B. Distributions of differences in confidence efficiency estimates between patients and controls. Horizontal colored bars indicate 95% confidence intervals. Same color description as panel A.

When quantifying metacognitive performance using the confidence efficiency measure of metacognition - which controls for first-order deficits - individuals with schizophrenia had similar confidence efficiency in the detection (−0.17 [-0.45, 0.06]), familiarity (−0.00 [-0.44, 0.31]) and recollection tasks (−0.10 [-0.58, 0.30]) (Figure 3B). Within each group, metacognitive performance was comparable across tasks (Controls: Visual detection - Familiarity: -0.11[-0.40, 0.20], Visual detection - Recollection: -0.18[-0.51, 0.17], Recollection - Familiarity 0.07[-0.26, 0.41]; Patients: Visual detection - Familiarity: -0.28[-0.55, 0.14], Visual detection - Recollection: -0.24[-0.57, 0.19], Recollection - Familiarity -0.03[-0.50, 0.42]).

Contrary to our predictions about domain-generality, metacognitive performance as measured with mixed-effects logistic regressions did correlate across tasks, neither did we find correlations with clinical traits such as positive, negative and disorganization syndromes and cognitive insight (see SI, Figure S7, 8).

Discussion

The present study aimed at characterizing metamemory and metaperception in people with schizophrenia while controlling for first-order deficits. In particular, we assessed metacognition in visual detection, familiarity, and recollection tasks. We hypothesized that people with schizophrenia would be specifically impaired in the meta-memory domain. At the first-order level, we found that people with schizophrenia had lower first-order performance in the three tasks compared to healthy controls, which confirms the importance of accounting for first-order deficits to quantify second-order processes specifically. When doing so, contrary to our hypothesis we found that metacognitive sensitivity was preserved among individuals with schizophrenia in the three tasks. In what follows, we discuss technical and conceptual aspects of our paradigm that should be considered to interpret this result, and then examine its clinical and theoretical significance.

A key contribution of this study is our attempt to match first-order performance between tasks for each participant using adaptive procedures, and between groups of participants using a generative model of confidence. We note that our adaptive procedure to match performance between tasks was successful when considering average performance, but not when considering task performance across levels of stimulus strength. In other words, we equated the overall performance but not the slopes between tasks in Figure 2A (see SI for details). A plausible explanation for this is a contextual effect. In session 1, blocks of visual detection trials were separated from blocks of memory trials, whereas in session 2 the three tasks were interleaved within each block of trials. Thus, the visual detection psychometric curve (SI, Figure S2c) from which we determined four visual contrast levels was obtained from a sequence of low-contrast perceptual stimuli (3 × 80 stimuli in a row), whereas during session 2 these low-contrast visual stimuli were interleaved with high-contrasted memory stimuli. This contextual effect might have resulted in a rightward shift (See Figure S3) of the visual detection psychometric curve, leading to underperformance in both groups in the visual detection task compared to the familiarity and recollection tasks.

Regarding between-groups task performance matching, early pilot versions of the present protocol aimed at equating memory performance between participants using adaptive staircases that manipulated either the number of encoding items, or the lag variable, but these attempts were not successful (no convergence). Instead, we accounted for differences in task-performance between groups by relying on measures of confidence efficiency from a recent generative model of confidence¹⁶, which enables the estimation of metacognitive abilities in factorial designs. Although this framework is recent and has not been fully benchmarked yet, we note that we found qualitatively similar results using a Bayesian logistic mixed-effects regression, which does not consider possible cognitive deficits but has the advantage of providing hierarchical estimates of metacognitive sensitivity, dealing with unbalanced data, and considering prior knowledge to compute Bayes factors. In contradiction to existing literature, both frameworks revealed no evidence for a metacognitive deficit in any of the three tasks. In fact, we found evidence for an absence of metacognitive deficit in memory tasks, and only inconclusive evidence in the perceptual domain. The absence of metacognitive deficit in schizophrenia was corroborated by an absence of difference regarding confidence biases. Indeed, contrary to several studies which did not control for first-order performance^30–32, we found no overconfidence in errors nor underconfidence in correct responses. One possibility is that the confidence biases previously reported in schizophrenia also stem from first-order deficits differences. Furthermore, contrary to previous behavioral results showing a positive link between false alarms and positive symptoms or proneness to hallucinations^33–35, our sample of patients had comparable rates of false alarms compared to healthy controls in the visual detection task. They committed more false-alarms in the memory tasks, interpreted as false recognitions, but no relationships were found between rates of false alarms and PANSS positive score (see SI).

At a conceptual level, the framing of our memory tasks in terms of familiarity and recollection processes may be questionable. Indeed, although our recognition memory tasks shared some features with usual familiarity and recollection tasks (in particular the testing questions which are respectively context-independent and context-dependent), there was no delay between encoding and testing phases, as we manipulated task difficulty with a variable lag. Therefore, one may consider our tasks to reflect working memory, which is also known to involve familiarity and recollection processes³⁶. To our knowledge, no study assessed metacognition related to short-term memory in schizophrenia. At first sight, our results seem to be in contradiction with the study by Berna et and colleagues⁸, which reported impaired metamemory in schizophrenia in a long-term (autobiographical) memory task, while controlling statistically for first-order performance. Yet, if our results are construed as evidence for preserved “short-term” metamemory in schizophrenia, the contradiction might be only apparent. A full taxonomy of metamemory processes is beyond the scope of the present study, and developing new paradigms to assess metacognitive performance in distinct subdomains of memory while controlling for first-order performance is one of the numerous challenges the metacognitive field is facing³⁷.

With these technical and conceptual considerations in mind, we can contextualize our findings and assess their clinical relevance. Our protocols focus on “in-the-moment” metacognition³⁸, i.e. confidence in trial-by-trial decisions, also known as “local” metacognition as opposed to more “global” evaluations^39–41. Metacognitive evaluations have been construed as hierarchically organized, where aggregated local judgments give rise to global self-beliefs about one’s performance within a cognitive task or domain⁴². Interestingly, it has been shown that global metacognitive evaluations can be altered independently from the local monitoring processes⁴³. Yet, as recently discussed⁴⁴, both local and global measures of metacognition may give an incomplete picture of metacognitive abilities from a clinical perspective. This concern is corroborated by the fact that our metacognitive measures are not correlated with several clinical dimensions of interest for schizophrenia (symptoms, cognitive insight, self-reported cognitive functioning see SI). Only perceptual reasoning assessed with WAIS matrix subtest scores were positively correlated with metacognitive performance, as reported previously⁴⁵. The need for paradigms that do justice to the breadth of the metacognition construct, i.e. including more cognitive domains, larger timescales, and theory of mind is now becoming acknowledged by the field.