The timing of confidence computations in human prefrontal cortex

Abstract

Knowing when confidence computations take place is critical for building a mechanistic understanding of the neural and computational bases of metacognition. Yet, even though a substantial amount of research has focused on revealing the neural correlates and computations underlying human confidence judgments, very little is known about the timing of confidence computations. To understand when confidence is computed, we delivered single pulses of transcranial magnetic stimulation (TMS) at different times after stimulus presentation while subjects judged the orientation of a briefly presented visual stimulus and provided a confidence rating. TMS was delivered to either the right dorsolateral prefrontal cortex (DLPFC) in the experimental group or to vertex in the control group. We found that TMS to right DLPFC, but not to vertex, led to increased confidence in the absence of changes to accuracy or metacognitive efficiency. Critically, equivalent levels of confidence increase occurred for TMS delivered between 200 and 500 ms after stimulus presentation. These results suggest that confidence computations occur during a broad window that begins before the perceptual decision has been fully made and thus provide important constraints for theories of confidence generation.

Introduction

Metacognition, the ability to assess the quality of our own decisions, is crucial for effective decision-making (Fleming et al., 2012; Koriat, 2007, 2007; Metcalfe & Shimamura, 1994; Nelson, 1990; Shimamura, 2000; Yeung & Summerfield, 2012). A substantial amount of research has focused on revealing the neural correlates underlying human confidence judgments (Fleming et al., 2012; Morales et al., 2018; Pereira et al., 2017; Shekhar & Rahnev, 2018; Shimamura, 2000; Yeon et al., 2020; Zheng et al., 2021). Although confidence computation is distributed across a network of brain regions (Morales et al., 2018; Yeon et al., 2020), many studies have pointed to an important role of the prefrontal cortex (Janowsky et al., 1989; Shimamura, 2000; Shimamura & Squire, 1986) and specifically the dorsolateral prefrontal cortex (DLPFC), which has been linked to confidence judgments (Fleming et al., 2012; Rounis et al., 2010; Shekhar & Rahnev, 2018).

However, although much progress has been made in discovering where confidence is computed in the brain, much less is known about the timing of confidence computation (Desender, Donner, et al., 2021; Dotan et al., 2018; Fetsch et al., 2014, 2018; Moran et al., 2015; Pleskac & Busemeyer, 2010). For example, while some models of confidence assume that confidence signals are present during the initial evidence accumulation stage (Dotan et al., 2018; Hellmann et al., 2022; Rahnev et al., 2016; Ratcliff & Starns, 2009; Vickers, 1979; Yu et al., 2015), other models postulate that confidence computation only begins after the decision is made (Herregods et al., 2023; Moran et al., 2015; Pleskac & Busemeyer, 2010). However, most studies in the literature are correlational, and thus cannot establish the critical window of confidence computation in the brain.

Recently, Shekhar and Rahnev (2018) used a causal manipulation that provides initial clues about the period of confidence computation. The authors delivered a train of three pulses of transcranial magnetic stimulation (TMS) to the right DLPFC at 250, 350, and 450 ms after stimulus onset in a perceptual decision-making task. They found that the TMS train of pulses decreased confidence compared to a control region (the primary somatosensory cortex) but could not determine exactly when confidence computation occurred besides the fact that some part of the window between 250 and 450 ms after stimulus onset is important.

To address more precisely the issue of when confidence computations occur, here we used single pulse TMS at four different times. Specifically, we delivered single pulses of TMS at 200, 300, 400, and 500 ms after stimulus onset and compared the results to TMS delivered simultaneously with stimulus onset (0 ms condition). Subjects judged the orientation of a briefly presented visual stimulus and reported their confidence. We delivered online TMS to the right DLPFC in the experimental group, and to vertex in the control group. We found that TMS to right DLPFC, but not to vertex, led to an increase in confidence without any changes to the accuracy or metacognitive efficiency. More importantly, the levels of confidence increase brought by TMS were the same across intervals between 200 and 500 ms after the stimulus presentation. These results suggest that confidence computations occur during a broad time window. Because the perceptual decision is unlikely to be made within 200 ms on a substantial proportion of trials, these results go against strong versions of the post-decisional theories of confidence where all confidence computations occur only after the decision has already been made.

Methods

Preregistration

We preregistered the sample size, exclusion criteria, and analyses for the DLPFC TMS group (https://osf.io/3ru2m). After the data for the DLPFC group were collected, we additionally collected data from a control group where we targeted vertex instead of DLPFC.

Subjects

A total of 76 subjects were enrolled in the study with 50 subjects in the experiment group (TMS to DLPFC) and 26 subjects in the control group (TMS to vertex). Based on our preregistered criteria, we excluded a total of 14 subjects. Specifically, we excluded 10 subjects who did not finish the experiment either because of TMS-related discomfort (seven subjects) or because they did not complete all trials before the end of the session (three subjects). Another subject was excluded because their data was lost because of a computer malfunction. Finally, we excluded three subjects for performance lower than 55% correct. Thus, the final sample size consisted of 62 subjects (22 females and 40 males) with 43 subjects in the experimental and 19 subjects in the control group. All subjects were right-handed, had a normal or corrected-to-normal vision, and had no history of seizure, family history of epilepsy, stroke, severe headache, or metal anywhere in the head. All subjects provided informed consent and were compensated $30 for two hours of total participation.

Task

Each trial began with subjects fixating on a small white dot (size = 0.05°) at the center of the screen for 500 ms, followed by a presentation of a Gabor patch (diameter = 3°) oriented either to the right (clockwise, 45°) or to the left (counterclockwise, 135°) of vertical for 100 ms. The Gabor patch was superimposed on a noisy background. Subjects indicated the orientation of the Gabor patch while simultaneously rating their confidence on a 4-point scale (where 1 corresponds to lowest confidence and 4 corresponds to highest confidence) via a single key press (Figure 1A). The four fingers of the left hand were mapped to the four confidence ratings for the left tilt response, whereas the four fingers of the right hand were mapped to the four confidence ratings for the right tilt response. For each hand, the index finger indicated a confidence of 1, whereas the pinky finger indicated a confidence of 4. The orientation of the stimulus (left/right) was chosen randomly on each trial.

Figure 1. Task.

(A) Trial sequence. Each trial began with a short fixation (500 ms), followed by the presentation of an oriented Gabor patch (100 ms). Subjects had to simultaneously indicate the tilt (left/right) of the Gabor patch and their confidence on a 4-point scale. (B) Timeline of TMS delivery. TMS was delivered as a single pulse 0, 200, 300, 400, or 500 ms after the stimulus onset. Subjects had a mean response time of 1078 ms.

We delivered online TMS as a single pulse on each trial at 0, 200, 300, 400, or 500 ms after the stimulus onset. We chose the 200–500 ms delays to coincide with the presumed time window of confidence computation. Indeed, in a previous study, we delivered TMS to DLPFC as a train of three pulses starting at 250 and ending at 450 ms after the stimulus onset and found an effect on confidence (Shekhar & Rahnev, 2018). Further, neuronal recordings from monkeys suggest that the discrimination response emerges about 200 ms after stimulus onset (Siegel et al., 2015), suggesting that confidence computation in human DLPFC is unlikely to happen much earlier than 200 ms. The 0-ms condition was chosen as a control against which to compare the four delay conditions.

Design and procedure

The main experiment consisted of four runs each consisting of five 40-trial blocks (for a total of 800 trials). The five possible TMS delays (0, 200, 300, 400, and 500 ms after the stimulus onset) were presented in a pseudorandom order such that within each group of five trials, each delay appeared once. We chose these delays because Shekhar and Rahnev (2018) had previously found that the period between 250 and 450 ms is important for confidence computations. We did not extend the delays further (to either shorter or longer delays), so that we could maximize the number of trials for each condition. The design and procedure were identical for the DLPFC and vertex groups except for the targeted site. Subjects were blind to the goal of the study or the exact location of stimulation.

At the beginning of the experiment, subjects underwent a behavioral training procedure without TMS. The training session started with a high Gabor patch contrast value (80%) and gradually progressed to lower contrast values (the last block had contrast values of 6%). Subjects were given trial-by-trial feedback on their performance during this training period. Then, subjects completed a 3-down-1-up staircase procedure (which results in approximately 79.4% accuracy; Macmillan & Creelman, 2005) to adaptively estimate the contrast for each subject. This staircase was conducted without feedback and yielded a mean contrast value of 6.64% (SD = 0.96%). We used the contrast value obtained for each subject for the rest of the experiment.

At the end of the training, subjects completed one practice TMS block using the contrast threshold estimated by the staircase procedure. The practice block was included to accustom subjects to receiving TMS while performing the task. The practice block was excluded from further analyses.

Apparatus

The stimuli were generated using Psychophysics Toolbox (RRID: SCR_002881) in MATLAB (The MathWorks, RRID:SCR_001622). During the training and the main experiment, subjects were seated in a dim room and were positioned 60 cm away from the computer screen (21.5-inch display, 1920 X 1080 pixel resolution, 60 Hz refresh rate).

Defining the targets for TMS targeting

We defined two sites as targets for TMS: right DLPFC and vertex. We used the vertex as the control site as is standardly done in previous research (Jung et al., 2016; Pitcher et al., 2008; Weissman-Fogel & Granovsky, 2019). Based on previous studies, we localized right DLPFC using the location of the F4 electrode in the 10–20 system used for EEG electrode placement (Conson et al., 2015; Fitzgerald, 2021; Fitzgerald et al., 2009; Mir-Moghtadaei et al., 2015; Rusjan et al., 2010). As in previous studies that targeted DLPFC with TMS during perceptual decision-making tasks (Rahnev et al., 2016; Shekhar & Rahnev, 2018), the DLPFC target was defined in the right hemisphere because the right hemisphere is dominant for visual processing (Hellige, 1996), although it is unclear whether the confidence computation itself is lateralized.

To determine the subject-specific location for stimulation, we located the F4 electrode location using skull measurements by following the algorithm developed by Beam et al. (2009). We did not localize right DLPFC using functional or anatomical data because such data were not available for our subjects. To compensate for the potentially lower accuracy of localization, we used a larger DLPFC subject sample than prior research on this topic (Rahnev et al., 2016; Rounis et al., 2010; Ryals et al., 2016; Shekhar & Rahnev, 2018). The location of the vertex was determined as the midpoint between the Nasion and inion.

TMS setup

TMS was delivered with a magnetic stimulator (MagPro R100; MagVenture, RRID:SCR_009601) using a figure-eight coil with a diameter of 75 mm. We determined the resting motor threshold (RMT) immediately before starting the main experiment. To localize the motor cortex, we marked its putative location and applied suprathreshold single pulses around that location. We determined the location of the right motor cortex as the region that induced maximal twitches of the fingers in the left hand. Then, using this location as the target, we determined the RMT using an adaptive parameter estimation by sequential testing procedure (Borckardt et al., 2006). For one subject, we were unable to estimate RMT reliably, so this subject was excluded from the experiment.

The TMS coil was oriented tangential to the skull such that the induced magnetic field was orthogonal to the skull. Stimulation was delivered at 120% of the individual RMT (average stimulation intensity = 72% of maximum stimulator output). In two cases, the stimulation intensity exceeded 80% of the maximum stimulator output. Due to discomfort, the intensity was reduced to 80% of the maximum stimulator output for both subjects. No arm or leg movements were elicited by stimulation of either DLPFC or vertex.

Analyses

We analyzed the accuracy, reaction time (RT), confidence, and metacognitive efficiency for each delay condition. Metacognitive efficiency was operationalized using the measure M-Ratio developed by Maniscalco and Lau (2012). M-Ratio is derived from signal detection theoretical modeling of the observer’s decision and confidence responses. It is the ratio of two measures: the observer’s metacognitive sensitivity (meta-d’, the ability to discriminate between correct and incorrect responses) and the observer’s stimulus sensitivity (d’, the ability to discriminate between the two stimulus classes). The ratio of meta-d’ to d’ factors out the contribution of stimulus sensitivity towards metacognitive performance and captures the efficiency of the observer’s metacognitive processes (Fleming & Lau, 2014).

To examine the effect of TMS, we computed the difference between confidence in each delay condition and confidence in the 0-ms condition. Then, we compared the obtained difference scores between two TMS stimulation sites (DLPFC and vertex) and the four TMS delay conditions (200, 300, 400, and 500 ms) using one-way and two-way repeated-measures ANOVAs. We then repeated the same procedure for M-Ratio instead of confidence. Direct comparisons between the two TMS stimulation sites within each delay condition were made using independent sample t-tests, whereas direct comparisons between different delay conditions within a single stimulation site were made using paired t-tests.

Note that the analyses performed above differ in some ways from the analyses we preregistered. No part of the study procedures or analysis plans for the control group were preregistered prior to the control group data collection. The reason for this is that our preregistration anticipated that there would be differences between the TMS effects for the four different delay conditions and that there may be individual variability between subjects as to the most effective delay condition. Because neither of these assumptions was supported by the data, the analyses we preregistered are subsumed within the simpler analyses we performed instead. Nevertheless, for completeness, we report the results of all preregistered analyses in the Supplementary Results.

Data and code availability

All data and code are available at https://osf.io/szr9u/. We report how we determined our sample size, all data exclusions, all inclusion/exclusion criteria, whether inclusion/exclusion criteria were established prior to data analysis, all manipulations, and all measures in the study.

Results

We used an online TMS protocol to investigate the timing of confidence computation by disrupting DLPFC activity at different time points (200, 300, 400, and 500 ms) after stimulus onset and compared the effects to a control condition where TMS was delivered over vertex. Subjects indicated the tilt (left/right) of a noisy Gabor patch while simultaneously providing a confidence rating on a 4-point scale. We compared the results to a condition where TMS was delivered at stimulus onset (i.e., 0 ms delay).

Previous work consistently found that TMS to DLPFC had no effect on either accuracy or reaction time (RT) (Rahnev et al., 2016; Rounis et al., 2010; Ryals et al., 2016; Shekhar & Rahnev, 2018). However, we observed that the 0-ms condition in the DLPFC group produced lower accuracy (74.1% correct) than the four delay conditions (200 ms: 76.5% correct, t(42) = 3.42, p = 0.001, Cohen’s d = 0.51; 300 ms: 75.5% correct, t(42) = 1.68, p = 0.04, Cohen’s d = 0.25; 400 ms: 76.2% correct, t(42) = 3.09, p = 0.002, Cohen’s d = 0.46; 500 ms: 76.1% correct, t(42) = 2.59, p = 0.007, Cohen’s d = 0.39). These results appear consistent with the notion that the 0 ms TMS may have induced an eye blink, drawn attention away from the stimulus, or otherwise interfered with the initial processing of the stimulus. Because the decrease in accuracy for the 0-ms condition occurred for some subjects only, for subsequent analyses we excluded all subjects for whom the 0-ms condition had accuracy more than 3.5% lower than the average of the four delay conditions. This led to the exclusion of 12 subjects in the DLPFC group while also equating the accuracy of the 0-ms condition (average accuracy = 75.69%) and average accuracy in the four delay conditions (average accuracy = 75.76%; t(30) = 0.18, p = 0.86, Cohen’s d = 0.007). We also applied the same exclusion criterion to the vertex group, which led to the exclusion of 3 subjects and also equated the average accuracy of the 0-ms condition (average accuracy = 75.98%) and the average accuracy in the four delay conditions (average accuracy = 76.11%; t(15) = −1.42, p = 0.17, Cohen’s d = −0.06). The lower rate of exclusion for the vertex condition is consistent with the possibility that TMS at 0 ms may have induced eye blinks for some subjects, but this happened primarily for DLPFC since that site is closer to the eye sockets. Repeating the analyses below without these exclusions still leads to the same main conclusions (Supplementary Figures 1–4).

We then examined the effects of TMS on task performance across the four delay conditions. A two-way repeated-measures ANOVA on the accuracy difference between each delay condition and the 0-ms baseline condition with factors TMS site (DLPFC and vertex) and delay conditions (200, 300, 400, and 500 ms) showed that there was no main effect of TMS site (F_(1,180) = 1.91, p = 0.13, η²= 0.01), no main effect of delay condition (F_(3,180) = 1.3, p = 0.13, η²= 0.03), and no interaction between TMS site and delay condition (F_(3,180) = 0.39, p = 0.76, η²= 0.01; Figure 2A). A similar two-way repeated-measures ANOVA on the RT difference between each delay condition and the 0-ms baseline condition also showed no effect of TMS site (F_(1,180) = 2.0, p = 0.16, η²= 0.01), delay condition (F_(3,180) = 0.24, p = 0.87, η²= 0.004), or an interaction between the two (F_(3,180) = 0.90, p = 0.90, η²= 0.003; Figure 2B). Pairwise comparisons between the DLPFC TMS and vertex TMS groups for each delay condition also showed no differences in either accuracy or RT (all p’s > 0.15, BF₀₁ values between 1.38 and 3.31). Similarly, averaging across all four delay conditions also produced no significant effects in either accuracy or RT for either the DLPFC or the vertex group (all p’s > 0.17, BF₀₁ values between 1.68 and 5.14). Thus, TMS had equivalent effects at delays between 200 and 500 ms for both accuracy and RT.

Figure 2. TMS effects on accuracy and RT.

(A) The effect of TMS on the accuracy difference between each delay condition and the 0-ms condition. The accuracy difference does not depend on either the TMS site, the delay condition, or the interaction between the TMS site and the delay condition. (B) The effect of TMS on RT difference between each delay condition and 0-ms condition. The RT difference does not depend on either the TMS site, the delay condition, or the interaction between the TMS site and the delay condition. Error bars represent SEM; n.s., p > .05.

Having established that the four delay conditions do not differentially affect performance, we examined whether TMS with different timing had a differential effect on confidence or metacognitive efficiency. We originally hypothesized that DLPFC TMS will lead to a decrease in confidence and that this effect will be stronger in some delay conditions than in others. However, the results showed opposite patterns for both of these hypotheses. First, instead of a decrease, TMS to DLPFC led to an increase in confidence for each delay condition compared to the 0-ms condition (200 ms: t(30) = 5.37, p = 8.30 X 10⁻⁶, Cohen’s d = 0.94; 300 ms: t(30) = 5.23, p = 1.20 X 10⁻⁵, Cohen’s d = 0.92; 400 ms: t(30) = 5.60, p = 4.29 X 10⁻⁶, Cohen’s d = 0.98; 500 ms: t(30) = 4.75, p = 4.75 X 10⁻⁵, Cohen’s d = 0.83; Figure 3). This effect was not present when TMS was delivered to the vertex (all p’s > 0.06, BF₀₁ values between 0.79 and 2.21). Further, all pairwise comparisons between DLPFC and vertex TMS showed significant differences in confidence (p < 0.001 for all four comparisons). Second, instead of the hypothesized differences among the four delay conditions, we found that the increase in confidence was equivalent for all four conditions. Indeed, a one-way ANOVA on the confidence in the 200–500 ms delay conditions for the DLPFC TMS group showed no significant effect of condition (F_(3,120) = 0.03, p = 0.99, η²= 7.30 X 10⁻⁴). A similar one-way ANOVA for the vertex group also showed no significant effect of condition (F_(3,60) = 0.28, p = 0.83, η²= 0.014). Direct comparisons between any pair of delay conditions for both the DLPFC TMS and vertex TMS groups confirmed the lack of any significant differences between the delay conditions (p > 0.09 for all 12 pairwise comparisons, BF₀₁ values between 2.46 and 5.20). Importantly, the confidence difference between the delay and 0-ms conditions did not correlate with the corresponding accuracy or RT differences (Supplementary Figure 5), suggesting that the confidence effects seen for the DLPFC group were independent of any potential TMS-related changes to the primary decision. These results demonstrate that the confidence computation does not occur within a narrow time window after the stimulus presentation.

Figure 3. TMS effects on confidence.

TMS to DLPFC increased confidence in each delay condition compared to the 0-ms baseline condition, whereas TMS to vertex did not affect confidence for any delay condition compared to the 0-ms baseline. Critically, the effects for both DLPFC and vertex TMS were equivalent across the four delay conditions. Error bars represent SEM; ***, p < .001.

Finally, we examined whether TMS affected metacognitive efficiency M-Ratio. We performed a two-way repeated-measures ANOVA on the M-Ratio difference between the delay conditions and the 0-ms condition with the TMS site (DLFPC vs. vertex) and delay condition (200, 300, 400, 500 ms) as factors. We found no main effect of TMS site (F_(1,180) = 1.75, p = 0.19, η²= 0.01), no main effect of delay condition (F_(1,180) = 0.77, p = 0.51, η²= 0.03), and no interaction between delay condition and TMS site (F_(3,180) = 0.62, p = 0.60, η²= 0.006). Pairwise comparisons between the DLPFC TMS and vertex TMS groups for each delay condition also showed no differences in M-Ratio difference scores (all four p’s > 0.06, BF₀₁ values between 0.76 and 3.31). These results demonstrate that, in line with previous findings (Shekhar & Rahnev, 2018), online TMS to DLPFC has no effect on metacognitive efficiency.

Discussion

Understanding the timing of the confidence computations is critical to uncovering the underlying mechanisms of human metacognition. However, despite much progress in other aspects of metacognitive judgments, exactly when confidence is computed is still unclear. To address this question, we tested how single-pulse TMS delivered to DLPFC between 200 and 500 ms after the stimulus onset affects confidence. We found that TMS to all four delay conditions significantly increased confidence in the DLPFC group, but not in a control group where TMS was delivered to the vertex. Importantly, there was no difference in the level of confidence increase among the different delay conditions. Our results demonstrate that confidence is computed over a relatively wide time interval that begins as early as 200 ms after stimulus onset.

Our findings provide evidence against strong versions of post-decisional models of confidence where all confidence-related computation is assumed to take place after a decision has already been made. This assumption is made by several prominent models. For example, the 2-stage dynamic signal detection (2DSD) model – perhaps the most influential model of confidence, choice, and RT – postulates an initial accumulation-to-bound stage that determines the decision, and a second confidence accumulation stage that determines the confidence rating (Pleskac & Busemeyer, 2010). Similarly, the collapsing confidence boundary (CCB) (Moran et al., 2015) and the recent model developed by Herregods et al. (2023) also postulate a similar 2-stage process where no confidence information appears to be computed before the initial decision has been made (Herregods et al., 2023). Note that these models postulate the same 2-stage process regardless of whether the primary decision and the confidence judgment are given simultaneously (as in the current study) or using separate button presses. Our results showed that TMS delivered as early as 200 ms after stimulus onset can change the confidence rating without affecting the stimulus sensitivity. Given that the average RT was over one second, the internal decision is likely to have been made in less than 200 ms on only a very small percentage of trials (if any). Therefore, our results suggest that decision and confidence processes are overlapped in time and thus challenge the models above given that they are inherently built on the assumption that no confidence-related processes occur before the decision has been made.

To be clear, our results do not question the existence of post-decisional processes that contribute to the confidence rating. There is ample empirical evidence that information presented after the decision has been made influences the resulting confidence rating, especially when confidence is given after the initial decision (Desender, Ridderinkhof, et al., 2021). Such findings parallel other literature that post-decisional evidence can lead to changes in the decision itself too (Resulaj et al., 2009). Our findings are perfectly consistent with the existence of post-decisional influences on confidence, but they are at odds with the idea that confidence is exclusively computed on signals arriving after the decision has been made.

Our findings are most consistent with theories that postulate that confidence is continuously computed in an online fashion starting from the initial stage of evidence accumulation (Dotan et al., 2018). For example, Dotan et al. (2018) employed a task where subjects continuously indicated their evolving decision using their finger and found that different finger kinematics (position vs. speed) reflected momentary decision and confidence variables independently of each other. A prolonged process of confidence evaluation that roughly overlaps with the decision process fits well with our findings that TMS delivered between 200 and 500 ms after stimulus onset has comparable effects on confidence judgments.

It should be noted that our finding that single-pulse DLPFC TMS delivered after the stimulus onset increased confidence goes in the opposite direction of the results of our previous study where a train of three pulses delivered to DLPFC decreased confidence (Shekhar & Rahnev, 2018). One possible explanation for these different results is that the single TMS pulse in the current study led to increased DLPFC activation, whereas the TMS train in the Shekhar and Rahnev study led to inhibition that resulted in decreased DLPFC activation (Caparelli et al., 2012; Romero et al., 2019). Indeed, it is indeed well known that different TMS parameters can lead to opposite neural and behavioral effects (Caparelli et al., 2012; Huang et al., 2005; Klomjai et al., 2015). This possibility also fits with our previous proposal that DLPFC reads out the strength of the sensory signal and relays it to aPFC, where the readout is translated into a confidence judgment after incorporating additional, non-perceptual factors (Shekhar & Rahnev, 2018). Thus, single-pulse TMS to DLPFC may have led to an excitation that amplified the confidence readout, whereas the train of pulses in Shekhar and Rahnev (2018) may have suppressed the readout. Nevertheless, we acknowledge that this explanation is speculative and is likely to be oversimplified. Yet, regardless of the underlying mechanisms, both studies support the notion that DLPFC is a critical node for confidence computation and also converge on the finding that online TMS to DLPFC does not affect metacognitive efficiency.

The current work has several limitations. First, the confidence computation is likely distributed across a network of brain regions (Morales et al., 2018; Yeon et al., 2020) but here we only targeted right DLPFC. Our results should not be interpreted as suggesting that confidence is localized to a specific brain area. It is an open question whether the timing of confidence computations elsewhere in the brain is different. Nevertheless, our results demonstrate that confidence computations are already ongoing at least somewhere in the brain by 200 ms after stimulus onset.

A second limitation is that, like in many TMS studies, there could be a concern about whether subjects intuited the purpose of the experiment and produced behavior accordingly. We believe that this is unlikely because (i) subjects were not informed about the purpose of the study, (ii) our own predictions were different from the results we observed (as can be seen from our preregistration), (iii) it would not be easy for subjects to perfectly identify the baseline 0-ms TMS given that they were concurrently engaged in a challenging perceptual task, and (iv) most importantly, the vertex and DLPFC groups underwent identical procedures but produced very different results. While subjects did have a tactile sensation about the approximate TMS location, most lacked the neuroscience background to know what processes the corresponding brain area is involved in. Therefore, we believe that it is unlikely that our results are due to factors outside of the direct neural effects of TMS.

A third limitation is that our results do not specify the precise computational mechanisms through which TMS affected confidence but not accuracy or RT. This question can in theory be addressed using computational modeling. However, popular models of choice, RT, and confidence such as the 2DSD (Pleskac & Busemeyer, 2010), balance-of-evidence (Vickers, 1979), RTCON (Ratcliff & Starns, 2013), and the recent drift-diffusion weighted-evidence-and-visibility model (Hellmann et al., 2023) make very different assumptions, and our data are not rich enough to distinguish between them. Thus, picking a specific way of modeling the current results may be arbitrary and any conclusion we draw would not necessarily be generalizable to other models.

A final limitation of our work is that TMS could have potentially disrupted not just online computation but also working memory-like representations. However, even if what is disrupted by TMS is a working memory-like signal, that would still support the conclusion that confidence computations occur very early (which challenges strong versions of post-decisional models of confidence). In addition, previous studies that used TMS to target the prefrontal cortex have sometimes shown exquisite temporal selectivity (Desrochers et al., 2015; Mottaghy et al., 2003; Muri et al., 1996). For example, Desrochers et al. (2015) found that TMS to RLPFC was only effective during a narrow period that extended for less than 80 ms, suggesting that what was disrupted was online computations rather than working memory-like representations. In line with such findings, we also believe that TMS in our study is likely to have disrupted online computations.

In conclusion, we found that single-pulse TMS to DLPFC delivered between 200 and 500 ms after stimulus onset increases confidence, but that a similar effect does not occur for vertex TMS. These results suggest that confidence computations take place during a broad time window in parallel with decision-making.

Figure 4. TMS effects on M-Ratio.

TMS to both DLPFC and vertex did not affect M-Ratio values when compared to the 0-ms baseline condition. There was also no significant effect of the TMS site, delay condition, or the interaction between the TMS site and delay condition. Error bars represent SEM; n.s., p > .05.