Pharmacological Elevation of Catecholamine Levels Improves Perceptual Decisions, But Not Metacognitive Insight

Descriptive statistics of RT, d’, and meta-d’

Participants discriminated the orientation of Gabor patches as clockwise (CW, 45°) or counterclockwise (CCW, −45°), while simultaneously reporting the confidence (high or low) in their perceptual judgment (Fig. 1B). To assess perceptual and metacognitive effects, we calculated perceptual sensitivity (d’), the overall ratio of high confident answers, and metacognitive sensitivity (meta-d’) derived from (extensions of) signal detection theory (SDT; Green and Swets, 1966; Fleming and Lau, 2014; see Materials and Methods). Furthermore, we calculated M-diff (meta-d’ − d’) and M-ratio (meta-d’ / d’) to assess whether drug condition affected the absolute (M-diff) and relative (M-ratio) difference between perceptual and metacognitive sensitivity. We used M-diff to perform a direct test of the interaction effect between drug condition and sensitivity type, necessary for establishing whether drug effects on perceptual sensitivity were larger or smaller than on metacognitive sensitivity in absolute terms (see Materials and Methods for further substantiation of this decision; Nieuwenhuis et al., 2011; Meyen et al., 2022).

Across all pharmacological conditions, mean reaction times (RT) were 690 ms (SD = 108 ms), and mean perceptual sensitivity (d’) was 1.46 (SD = 0.48; corresponding to ∼75% correct, i.e., titrated level; see Materials and Methods). Participants had metacognitive insight in their performance, as meta-d’ was significantly above zero (µ meta-d’ = 0.75, SD = 0.40; one-sample t test against 0: t₍₂₆₎ = 9.90, p < 0.001, d = 1.91). Overall, metacognitive evaluations were “inefficient,” however, because M-diff was significantly lower than 0 (µ M-diff = −0.70, SD = 0.50; one-sample t test against 0: t₍₂₆₎ = −7.27, p < 0.001, d = 1.40), as observed more often in similar tasks (Hauser et al., 2017; Samaha and Postle, 2017; Shekhar and Rahnev, 2021a,b). A subzero M-diff is often interpreted as an indication that less sensory evidence was available for the metacognitive decision as compared with the perceptual decision (Fleming and Lau, 2014).

Catecholaminergic enhancement improves perceptual, but not metacognitive decision-making

Previously, we established that ATX administration enhances d’ (no effect of DNP; Nuiten et al., 2023), and here we replicate this observation: ATX increased d’ compared with PLC, whereas DNP did not (ATX, one-sided t test: t₍₂₆₎ = 2.06, p = 0.02, d = 0.39, BF₀₁= 0.40; DNP: t₍₂₆₎ = 1.04, p = 0.31, d = 0.20, BF₀₁= 3.00; Fig. 1C). Drug condition did not significantly affect the ratio of high confidence answers (ATX: t₍₂₆₎ = 0.49, p = 0.63, d = 0.09, BF₀₁= 4.39; DNP: t₍₂₆₎ = 0.04, p = 0.97, d = 0.01, BF₀₁= 4.90; Fig. 1D). Contrary to the effects on d’, neither ATX nor DNP significantly modulated meta-d’ and Bayesian analyses provided moderate evidence for the absence of effects (ATX: t₍₂₆₎=−0.84, p = 0.41, d = 0.16, BF₀₁= 3.55; DNP: t₍₂₆₎=−0.46, p = 0.65, d = 0.07, BF₀₁= 4.44; Fig. 1E). Next, we further substantiated whether the effects of ATX on d’ (enhancement) and meta-d’ (no effect) indeed differed, by testing the interaction between drug condition (ATX/DNP vs PLC) and measurement (d’ vs meta-d’; Nieuwenhuis et al., 2011; Meyen et al., 2022). To do so, we calculated M-diff (meta-d’ − d’) and show that it was lower for ATX versus PLC, but not for DNP versus PLC (ATX: t₍₂₆₎=−2.13, p = 0.04, d = 0.47, BF₀₁= 0.71; DNP: t₍₂₆₎=−1.21, p = 0.24, d = 0.23, BF₀₁= 2.54; Fig. 1F). Thus, ATX specifically improved perceptual decision-making (d’) without affecting metacognitive decision-making (meta-d’) and significantly increased the absolute difference between meta-d’ and d’ (M-diff), showing that ATX disproportionally increased perceptual over metacognitive decision-making.

Additionally, we analyzed whether the ratio between d’ and meta-d’, reflected in the measure M-ratio (meta-d’ / d’) was altered by our pharmacological manipulation. We used a hierarchical Bayesian approach to obtain posterior distributions of log-transformed M-ratio estimates, for each of the three drug conditions (see Materials and Methods). We calculated the difference between the posterior distribution of the active drug condition (ATX and DNP) and the posterior distribution of PLC (Fig. 1G). Although for both ATX and DNP, the mean difference in M-ratio with PLC was below zero, indicative of a lowered M-ratio, the 95% credibility intervals (CI_95%) overlapped with zero suggesting that M-ratio was likely not affected by either drug (difference ATX-PLC log(M-ratio), μ = −0.25, SD = 0.16, CI_95%=[−0.59, 0.07]; difference DNP-PLC log(M-ratio), μ = −0.09, SD = 0.13, 95% CI_95%=[−0.36, 0.17]). Indeed for both drugs, the probability that the drug effect was larger than 0 was larger than 0.05, suggesting that neither drug modulated M-ratio [ATX: p(difference > 0) = 0.06; DNP: p(difference > 0) = 0.23; Fig. 1G].

Due to repeated exposure to the task, and fixed task difficulty across drug sessions, d’ significantly increased over the course of the three sessions (one-factor repeated-measures ANOVA; F_(2,52) = 10.97; p < 0.001; η_p²= 0.30; BF₀₁= 0.01). Training in the behavioral task did not, however, result in a change in meta-d’ (one-factor repeated-measures ANOVA; F_(2,52) = 1.30; p = 0.28; η_p²= 0.05; BF₀₁= 3.43). Although the order of drug administration was counterbalanced between individuals, we wanted to exclude the possibility that the main effects of ATX on d’ were driven by training effects. To do so, we performed a mixed-level ANOVA, in which we included drug condition as the within-subject variable and drug order (consisting of six levels, given there were three drugs) as between-subject variable. We only performed this analysis for ATX (vs PLC), as DNP did not influence d’ whatsoever. The mixed-level ANOVA revealed no main effect of drug order (F_(5,21) = 0.74; p = 0.60; η_p²= 0.15; BF₀₁= 4.18) but did show a main effect of drug condition (F_(1,21) = 4.66; p = 0.04; η_p²= 0.18; BF₀₁= 0.69). Crucially, there was no interaction between drug order and drug condition (F_(5,21) = 1.50; p = 0.23; η_p²= 0.26; BF₀₁= 0.84). Therefore, it is unlikely that the effect of ATX on d’ was driven by behavioral training of the task.

Neither ATX nor DNP significantly affected reaction times (RT; ATX: t₍₂₆₎ = 0.82, p = 0.42, d = 0.09, BF₀₁= 3.62; DNP: t₍₂₆₎ = 1.00, p = 0.33, d = 0.13, BF₀₁= 3.13), response criterion (c; ATX: t₍₂₆₎ = 0.12, p = 0.91, d = 0.02, BF₀₁= 4.88; DNP: t₍₂₆₎ = 0.45, p = 0.66, d = 0.08, BF₀₁= 4.48), and metacognitive bias (meta-c; ATX: t₍₂₆₎ = 0.31, p = 0.76, d = 0.05, BF₀₁= 4.69; DNP: t₍₂₆₎ = 0.05, p = 0.96, d = 0.00, BF₀₁= 4.90) as compared with PLC.

Summarizing, ATX enhanced perceptual sensitivity (d’) but did not significantly affect metacognitive sensitivity (meta-d’). Moreover, ATX significantly lowered M-diff, demonstrating that it disproportionally increased the absolute difference between perceptual sensitivity and metacognitive sensitivity. Although the direction of the ATX effect on M-ratio was in line with its effect on M-diff, the 95% credibility interval for this effect overlapped with zero suggesting the absence of an effect of ATX on M-ratio. Elevated cholinergic levels did not robustly modulate behavior in any way.

Prestimulus pupil-linked arousal partially mimics ATX effects

Thus, the pharmacological elevation of catecholaminergic activity (through ATX) affected perceptual sensitivity in the absence of metacognitive sensitivity effects. To further explore the reliability and generality of these effects, we next focused on a measure thought to reflect spontaneous or task-related fluctuations in, among others, catecholaminergic (but also cholinergic) activity, namely, pupil size (Reimer et al., 2014; McGinley et al., 2015; Joshi et al., 2016). Fluctuations in prestimulus pupil-linked arousal have previously been shown to modulate d’ during perceptual decision-making tasks (McGinley et al., 2015; Waschke et al., 2019; Podvalny et al., 2021; Beerendonk et al., 2024), but its relation to meta-d’ and M-diff has not been established. To assess the relation between fluctuations in pupil-linked arousal on our measures of task performance, we subdivided all trials into five equally sized bins based on the average pupil size in a short window (−500 to 0 ms) before the presentation of each Gabor stimulus (similar to Beerendonk et al., 2024; see Materials and Methods). We performed the binning procedure within each drug session and for each of the two experimental blocks separately. This approach ensures that each of the five pupil bins contains the same number of trials from each drug condition and experimental block, to exclude the possibility that pupil bin number would be driven by pharmacological effects or slow drifts in pupil size. Next, d’, overall confidence rate, meta-d’, and M-diff were calculated for every participant and pupil bin. Because the relationship between pupil size and choice behavior has been shown to be inverted u-shaped as well as linear, first- and second-order polynomial regressions were fitted (see Materials and Methods).

Replicating previous work (McGinley et al., 2015; Beerendonk et al., 2024), and in line with the behavioral results associated with our drug manipulation, d’ exhibited a clear relationship with prestimulus pupil diameter (Fig. 1H, top left panel). Specifically, both first-order (β₁) and second-order (β₂) fits were significant showing that d’ increased with pupil diameter for low to medium levels of pupil-linked arousal but was lowered at the extremes of high arousal (β₁: t₍₂₆₎ = 3.79, p < 0.001, d = 0.73, BF₀₁= 0.02; β₂: t₍₂₆₎ = −3.79, p < 0.001, d = 0.73, BF₀₁= 0.02). This increase in d’ was matched in decision confidence, which was also enhanced with elevated levels of pupil-linked arousal (β₁: t₍₂₆₎ = 2.89, p = 0.01, d = 0.56, BF₀₁= 0.17; β₂: t₍₂₆₎ = −2.66, p = 0.01, d = 0.51, BF₀₁= 0.27; Fig. 1H, top right panel). Prestimulus pupil-linked arousal was not related to meta-d’ (β₁: t₍₂₆₎ = 0.56, p = 0.58, d = 0.11, BF₀₁= 4.26; β₂: t₍₂₆₎ = −1.00, p = 0.33, d = 0.19, BF₀₁= 3.13; Fig. 1H, bottom left panel). Crucially, baseline pupil-linked arousal exhibited a negative linear relation with M-diff, indicating that d’ and meta-d’ were increasingly disproportionally affected for higher levels of pupil-linked arousal (β₁: t₍₂₆₎ = −3.12, p = 0.004, d = 0.60, BF₀₁= 0.11; β₂: t₍₂₆₎ = 1.07, p = 0.29, d = 0.21, BF₀₁= 2.92; Fig. 1H, bottom right panel). Finally, we tested for a similar linear relation between prestimulus pupil bin and M-ratio by fitting a hierarchical Bayesian model (see Materials and Methods). Although the mean of the posterior distribution fell below zero (μ = −0.02, SD = 0.03; in line with a negative linear relation between pupil bin and M-ratio), the 95% CI of the posterior overlapped with 0 (CI_95% = [−0.08, 0.04]; Fig. 1I).

Thus, similar to pharmacological elevations of catecholaminergic activity, elevated prestimulus pupil size—indicative of global arousal changes and driven by among other catecholamine activity—was associated with changes in perceptual sensitivity but not metacognitive sensitivity. Furthermore, the absolute, but not relative, difference between perceptual and metacognitive sensitivity increased with elevated prestimulus pupil size.

Catecholamines do not result in sharpened encoding of visual information

Next we employed the EEG data to investigate whether electrophysiological responses were modulated by drug, decision accuracy, and confidence. Note that neural effects related to metacognitive insight should manifest itself, if present, in the interaction between accuracy and confidence. Specifically, metacognitively correct answers are defined as high confidence correct and low confidence incorrect decisions, whereas metacognitively incorrect answers are defined as low confidence correct and high confidence incorrect decisions (Fleming and Lau, 2014).

We performed MVPA to test how the encoding of the task-relevant visual stimulus feature (Gabor orientation) was affected by drug administration and how it related to decision accuracy, confidence, and metacognition. To this end, we trained a linear discriminant classifier (LDA) on data from an independent task (referred to as the localizer task) to discriminate CW versus CCW stimuli. By training the classifier on this independent dataset, we ensured that it was not being contaminated by motor preparation/execution (see Materials and Methods; Fig. 2A). During this localizer task, Gabor patches were presented at a slight tilt (−2° or +2°) rendering their orientation as being closer to the vertical axis (−43° or 43°) or the horizontal axis (−47° or 47°). Participants reported Gabor orientation as being closer to the vertical or the horizontal axis via button presses (see Materials and Methods). Note that during the localizer task, Gabor patches were presented at full opacity.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Multivariate decoding of Gabor orientation. A, During the localizer task (illustration shown), participants reported a slight tilt offset of ±2° from diagonal, by pressing a button with their left hand (for horizontally tilted Gabors: −47° and +47°) or right hand (for vertically tilted Gabors: −43° and +43°). Gabor orientation and behavioral responses were therefore independent (see Materials and Methods for details). B, Temporal generalization of stimulus orientation decoding in the localizer task (AUC). Training time is shown on the y-axis, testing time on the x-axis. Black demarked areas represent clusters that survived cluster-based permutation testing at p ≤ 0.05. C, Topographic maps showing the activity patterns (from transformed classifier weights) in two separate time windows. D, Tenfold within-localizer decoding AUC taken from the cluster of panel B, separated by drug. E, Schematic of cross-task decoding setup. A linear discriminant was trained to distinguish CCW (red dots, example data) from CW (blue dots) during the localizer task (top panel). This linear discriminant was then presented with data from the main task, after which it classified trials as containing CCW or CW information (bottom panel). F, Temporal generalization of stimulus orientation cross-decoding (AUC) when training on the (localizer task, y-axis) and testing on the main task (discrimination task, on the x-axis). Black demarked areas represent clusters that survived cluster-based permutation testing at p ≤ 0.05. G, Cross-decoding AUC taken from the cluster of panel E, separated by drug, accuracy, and confidence. Note that conf. demarks the factor decision confidence and acc. demarks the factor decision accuracy. All error bars indicate SEM.

We first quantified how Gabor orientation was represented in neural data of the localizer task by performing a 10-fold within-localizer decoding procedure using a temporal generalization approach (GAT; Fig. 2B). Orientation representations in neural activity were observed on the diagonal of the GAT matrix from 70 to 640 ms poststimulus, in line with previous studies (Fig. 2B; King et al., 2016). There was limited off-diagonal decoding, suggesting that orientation representations were “chained” and constantly changed over time during the localizer task (King and Dehaene, 2014). Such rapidly changing neural representations have previously been observed for response-irrelevant stimulus features, such as Gabor orientation in the localizer task (Mostert et al., 2016; King and Wyart, 2021). Activation patterns were retrieved from decoding weights (as raw classifier weights are not directly interpretable as reflecting neural activations; Haufe et al., 2014) and showed that orientation information was initially present over occipital electrodes and was later most prominent over centroparietal ones (Fig. 2C). Drug condition did not modulate decoding accuracies within the cluster reported in Figure 2B, and Bayes factors revealed moderate evidence for an absence of any drug effect (ATX: t₍₂₆₎ = −0.91, p = 0.37, d = 0.22, BF₀₁= 3.37; DNP: t₍₂₆₎ = −0.66, p = 0.52, d = 0.16, BF₀₁= 4.03; Fig. 2D).

After verifying that (task-irrelevant) Gabor orientation information could be decoded from neural data of the localizer task, we next performed a cross-task decoding analysis in which classifiers that were trained on the localizer and tested on all timepoints of the main task (for more details see Materials and Methods). We first tested the classifier on all data from the discrimination task (irrespective of decision accuracy and confidence) to test whether we could cross-decode Gabor orientation (Fig. 2E). When averaging across drug sessions, we observed that stimulus orientation could be cross-decoded in a cluster from ∼430 to 945 ms (p = 0.002, cluster-corrected; Fig. 2F). The observed cluster was slightly off-diagonal, suggesting that there was a delay in the times when orientation information was initially decodable from the localizer task as compared with the discrimination task. This is likely because Gabor stimuli were presented at full contrast during the localizer task and at low contrast in the main task and increased stimulus contrast is known to speed up the latency of visually evoked responses (Groen et al., 2022). Neural representations generalized more over time when compared with the within-localizer decoding, possibly because orientation information during the main task was directly response relevant and thus needed to be sustained over time (King and Dehaene, 2014).

Next, we tested how drug condition, decision accuracy, decision confidence, and their interactions were related to sensory representations of Gabor orientation. To do so, we tested the classifier on EEG data from the discrimination task in a 3 × 2 × 2 (drug × accuracy × confidence) factorial design; i.e., the classifier was tested separately on the data of each of the 12 subconditions. This procedure effectively treats decision accuracy and confidence as independent variables, similar to the pharmacological manipulation. Although it is generally inappropriate to combine causal and correlational effects, it is necessary when investigating how neural processes leading up to specific behavioral outputs (e.g., errors) are modulated by an independent variable (e.g., pharmacological manipulation). Indeed, such approaches are common in research investigating the neural mechanisms underlying, for example, error processing, e.g., by testing causal effects on the error-related negativity (ERN) component (Tieges et al., 2004; Zirnheld et al., 2004; Barnes et al., 2014).

After testing the classifier on data from the main task, we extracted AUC values from the cluster reported in Figure 2F for every participant and subcondition and then performed a 3 × 2 × 2 (drug × accuracy × confidence) factorial rmANOVA to establish effects of accuracy, decision confidence and their interaction on classifier AUC (ignoring omnibus drug effects, while leveraging the full data design), and pairwise 2 × 2 × 2 (drug × accuracy × confidence) factorial rmANOVAs to test main and interaction effects of ATX and DNP versus PLC. We observed increased decoding AUC for correct versus incorrect answers (F_(1,26) = 26.85; p < 0.001; η_p²= 0.51; BF₀₁= 1.14 × 10⁻³; Fig. 2G), but decision confidence (F_(1,26) = 0.36; p = 0.55, η_p²= 0.01; BF₀₁= 5.13; Fig. 2G) and metacognitive accuracy were not associated with changes in AUC (interaction accuracy × confidence: F_(1,26) = 0.89; p = 0.35; η_p²= 0.03; BF₀₁= 3.39; Fig. 2G). In line with previous work (Alilović et al., 2023), decoding AUC was significantly below chance level for incorrect decisions (one-sample, two-sided t test against 0.50: t₍₂₆₎ = −4.34; p < 0.001; d = 0.83; BF₀₁= 6.83 × 10⁻³), suggesting that neural activity of incorrect trials contained “illusory” stimulus orientation information, i.e., information about the nonpresented stimulus orientation that was reported. Neither ATX nor DNP modulated overall decoding accuracy (ATX: F_(1,26) = 0.01, p = 0.92, η_p²= 0.00, BF₀₁= 4.91; DNP: F_(1,26) = 0.02, p = 0.88, η_p²= 0.00, BF₀₁= 4.94). Furthermore, neither drug significantly modulated how decision accuracy, decision confidence, and metacognitive accuracy (three-way interaction) were reflected in classification AUC (all p > 0.48).

Summarizing, although Gabor orientation could be reliably decoded from posterior and central electrodes, drug condition did not affect orientation decoding in any way, neither in the localizer task nor in the main task, and there were no interactions with decision accuracy and confidence in the latter. This suggests that elevated cholinergic and catecholaminergic levels do not modulate the encoding of task-relevant visual stimulus features and that the behavioral effects of catecholaminergic enhancement (on d’ and M-diff) were likely not related to changes in the encoding of task-relevant stimulus features.

Catecholaminergic modulations of centroparietal evidence accumulation signals

The process of perceptual decision-making is thought to (sequentially) rely on the extraction of sensory information from stimulus input (reported above) followed by the integration of that sensory information into a (unsigned, i.e., not containing stimulus category information) decision variable that accumulates over time until a certain internal decision threshold is crossed (Gold and Shadlen, 2007; Ratcliff and McKoon, 2008; Shadlen and Kiani, 2013; Msheik et al., 2022). Previous primate work has shown that the parietal cortex is key for integrating sensory information into this decision variable (Gold and Shadlen, 2007; Kiani and Shadlen, 2009) and in humans, the CPP has been identified as an ERP component thought to reflect sensory evidence accumulation (O’Connell et al., 2012; Kelly and O’Connell, 2013; Loughnane et al., 2016; van Kempen et al., 2019; van Vugt et al., 2019; Nuiten et al., 2023). The CPP scales with RT, decision accuracy, and decision confidence (O’Connell et al., 2012; Kelly and O’Connell, 2013; van Vugt et al., 2019; Nuiten et al., 2023), and its peak amplitude has been shown to increase with elevated levels of catecholamines but decrease with elevated levels of acetylcholine (Loughnane et al., 2019; Nuiten et al., 2023).

In Figure 3B, we plotted the response-locked CPP for ATX, DNP, and PLC, for high versus low confidence trials and correct versus incorrect trials average across three electrodes over centroparietal regions [see Fig. 3A, bottom topomap for spatial region of interest (ROI) used for this analysis]. Electrode selection was identical to our previous study (Nuiten et al., 2023) and in line with other work in this field (O’Connell et al., 2012; Kelly and O’Connell, 2013; Loughnane et al., 2016; van Kempen et al., 2019). We replicated our previous finding that ATX increased CPP amplitudes, here in a time window from −180 to 0 ms preceding the response (cluster-corrected p = 0.04; Fig. 3B). Moreover, we observed increased CPP amplitudes for correct versus incorrect trials (cluster-corrected p = 0.001; −211 to 0 ms; Fig. 3B) and high confident versus low confident trials (cluster-corrected p = 0.01; −117 to 0 ms; Fig. 3B). None of the other effects survived cluster-correction or did we observe significant effects before −600 ms preresponse. In Figure 3C, we plot the mean CPP amplitudes for the full 3 × 2 × 2 design, averaged across the temporal cluster in which the main effect of ATX was significant (−180 to 0 ms). In sum, we found that ATX, decision accuracy, and confidence were all associated with increased CPP amplitude, but metacognitive accuracy was not reflected in the CPP (also not in interaction with ATX). ATX thus mainly increased the overall amplitude of the CPP but did not modulate the relationship between accuracy (correct/incorrect) and confidence (high/low). We did not replicate our previous observation that DNP attenuates the peak of the CPP (Nuiten et al., 2023).

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

Response-locked ERPs over centroparietal and frontal cortical regions. A, Topographic maps of response-locked activity, in steps of 100 ms. The bottom topographic map shows spatial ROIs for the CPP (three electrodes at over centroparietal areas: CP1, CP2, CPz) and the FP (three electrodes over frontal areas: F1, F2, Fz) analyses as white markers. B, Response-locked CPP shown for the ATX, DNP, and PLC condition, split up for high confident (solid lines) versus low confident (dotted lines) and correct (opaque lines) versus incorrect (transparent lines) trials. The horizontally plotted black dots indicate effects that survived across time cluster-correction. *The effects of ATX were tested with a cluster-corrected one-sided t test (ATX > PLC, see Materials and Methods). C, Average CPP amplitudes from the cluster in which ATX modulated CPP amplitude. The three drug conditions are plotted on the x-axis, separated for high and low confidence decisions. Correct trials are plotted in opaque lines and incorrect trials in transparent lines. The inset bar plot on the right shows the overall CPP amplitude (averaged across confidence and accuracy) for each drug. D, Same as B, but for FP ERP component. E, Average FP amplitudes within the cluster where the three-way interaction (drug × accuracy × confidence) was significant. The three drug conditions are plotted on the x-axis, separated for high and low confidence decisions. Correct trials are plotted in opaque lines and incorrect trials in transparent lines. The inset bar plot on the right shows the FP amplitude interaction effect (accuracy × confidence) for each drug, defined as [(C_hi − C_lo) + (I_lo − I_hi)], with C_hi, correct decision, high confidence; C_lo, correct decision, low confidence; I_lo, incorrect decision, low confidence; and I_hi, incorrect decision, high confidence. All error bars indicate SEM.

Catecholaminergic modulations of frontal EEG signals reflecting metacognitive insight

Recent fMRI studies have revealed that the computation of metacognitive insight during task performance is associated with activity in human prefrontal cortex (PFC) and causal interference of the anterior part of this region using TMS confirmed its crucial role in metacognition (Fleming et al., 2012b; D. Bang and Fleming, 2018; Morales et al., 2018; Lapate et al., 2020). Furthermore, human electrophysiology has revealed a (response-locked) ERP component averaged over frontal electrodes (F1, F2, Fz; see Fig. 3A, bottom row for spatial ROI), referred to as the frontal potential (FP; Lim et al., 2020; Feuerriegel et al., 2022), which has been shown to reflect upcoming decision confidence (Lim et al., 2020), but only for correct trials (Feuerriegel et al., 2022). This suggests that it may reflect processes associated with computing metacognitive evaluations of task performance.

In Figure 3D we plot the response-locked FP for all 12 conditions for the frontal ROI. Electrode selection was based on previous studies focusing on the FP component (Lim et al., 2020; Feuerriegel et al., 2022). In general, the FP component showed a gradually declining profile toward response execution, becoming more negative over time. Interestingly, we observed that ATX decreased FP amplitude even further as compared with PLC from −594 to −125 ms before the button press (cluster-corrected p = 0.002; Fig. 3D). DNP did similarly, but for a shorter time period (−586 to −320 ms, cluster-corrected p = 0.01). More importantly, we observed a three-way interaction between ATX (vs PLC), decision accuracy, and confidence right before response execution (from −234 to −55 ms, cluster-corrected p = 0.02; Fig. 3D). In Figure 3E, we plot FP amplitudes per drug condition separately to further reveal its profile. As can be observed, in the PLC and DNP condition, the amplitude of the FP clearly shows the interactive pattern indicative that FP tracks metacognitive accuracy. Specifically, the FP was least negative for correct answers given with high confidence, whereas FP amplitude decreased for correct answers with low confidence, replicating previous work (Feuerriegel et al., 2022). For incorrect answers, the confidence modulation was more subtle. Metacognitively correct answers (low confidence incorrect answers) were associated with the least negative FP (especially in the PLC condition), whereas metacognitively incorrect answers (high confidence incorrect answers) were associated with a relatively more negative FP. This is in line with the idea that the FP reflects metacognitive insight, most prominently in correct trials (Lim et al., 2020; Feuerriegel et al., 2022). However, remarkably, the ordering of these four conditions changed drastically when ATX was administered. To further quantify the effect of ATX, we calculated the overall “metacognitive effect” for each drug (Fig. 3E, right panel), reflecting the difference in FP amplitude between metacognitively correct answers and metacognitively incorrect answers ([high confidence correct − low confidence correct trials] + [low confidence incorrect − high confidence incorrect trials]). This measure can be interpreted as a reflection of overall metacognitive performance in neural activity. We observed that ATX inverted the magnitude of this effect entirely compared with PLC. This condition clearly indicates a lack of association between the FP and metacognitive insight, in line with the idea that ATX impoverishes metacognitive insight, compared with perceptual task performance (Fig. 1). In the Discussion, we will further elaborate on this finding.

Pupil-linked arousal is related to a metacognitive signature in the FP

In a final analysis, similar to Figure 1, we used prestimulus pupil size as an indirect index of neuromodulator activity, rather than the pharmacological manipulation, and tested its relation to the univariate EEG effects. In Figure 4, we plot the CPP and FP for each of the five pupil bins, separated by decision accuracy and confidence. We calculated the mean CPP and FP amplitudes in the same time windows as reported in Figure 3C (−180 to 0 ms) and Figure 3E (−234 to −55 ms), but now per pupil bin instead of drug condition. For the CPP, we tested the presence of a linear relation with pupil bin, to directly link this analysis to the observed main effect of ATX on the CPP (Fig. 3C). Regression analysis did not reveal a significant relation between CPP amplitude and prestimulus pupil size, although numerically the effect went in the expected direction (β₁: t₍₂₆₎ = 1.41, p = 0.17, d = 0.27, BF₀₁= 2.02; β₂: t₍₂₆₎ = −0.57, p = 0.58, d = 0.11, BF₀₁= 4.24; Fig. 4B). Next, we extracted FP amplitudes from the time window in which we observed the cluster-corrected three-way interaction effect (ATX × accuracy × confidence; Fig. 3D). Then, we calculated the average interaction effect between decision accuracy and confidence (similar to above) within these timepoints and tested its relation with prestimulus pupil size. Regression analysis revealed that the metacognitive signature in the FP (explained above) became more negative with increasing prestimulus pupil size (β₁: t₍₂₆₎ = 2.62, p = 0.01, d = 0.51, BF₀₁= 0.29; β₂: t₍₂₆₎ = −0.95, p = 0.35, d = 0.18, BF₀₁= 3.27; Fig. 4D), suggesting that metacognitive accuracy, as reflected in FP, deteriorated when going from states of low versus high pupil-linked arousal.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

Relation between pupil size and the CPP and FP. Same as Figure 3, but now using five prestimulus pupil size bins as variables instead of the pharmacological manipulation. A, Response-locked CPP shown for each pupil size bin, split up for high confident (solid lines) versus low confident (dotted lines) and correct (opaque lines) versus incorrect (transparent lines) trials. B, CPP amplitudes extracted from the temporal cluster in A (statistics and cluster from Fig. 3B), shown for each condition as a factor of pupil size. The right panel shows the CPP amplitude per pupil bin and the regression results. C, Response-locked FP (accuracy × confidence), as a factor of pupil size bin. D, FP amplitudes extracted from cluster in C (same as in Fig. 3D), shown for each condition. The inset bar plot on the right shows the FP amplitude interaction effect (accuracy × confidence) for each drug, defined as [(C_hi − C_lo) + (I_lo − I_hi)], with C_hi, correct decision, high confidence; C_lo, correct decision, low confidence, high confidence; I_lo, incorrect decision, low confidence; and I_hi, incorrect decision. Note that the regression analyses of the utmost right panels were performed with bin number as independent variable, but the x-axis shows the mean pupil size (in % of block mean) for each bin. All error bars indicate SEM.