Interpersonal alignment of neural evidence accumulation to social exchange of confidence

  1. Jamal Esmaily  Is a corresponding author
  2. Sajjad Zabbah
  3. Reza Ebrahimpour  Is a corresponding author
  4. Bahador Bahrami  Is a corresponding author
  1. Department of General Psychology and Education, Ludwig Maximillian University, Germany
  2. Faculty of Computer Engineering, Shahid Rajaee Teacher Training University, Islamic Republic of Iran
  3. Graduate School of Systemic Neurosciences, Ludwig Maximilian University Munich, Germany
  4. School of Cognitive Sciences, Institute for Research in Fundamental Sciences (IPM), Islamic Republic of Iran
  5. Wellcome Centre for Human Neuroimaging, University College London, United Kingdom
  6. Max Planck UCL Centre for Computational Psychiatry and Aging Research, University College London, United Kingdom
  7. Institute for Convergent Science and Technology, Sharif University of Technology, Islamic Republic of Iran
  8. Centre for Adaptive Rationality, Max Planck Institute for Human Development, Germany
5 figures, 8 tables and 2 additional files

Figures

Figure 1 with 4 supplements
Experiment paradigm and behavioral results.

(a) Timeline of trials in isolated (top) and social (bottom) conditions. After stimulus presentation, subjects reported their decision and confidence simultaneously by clicking on 1 of the 12 …

Figure 1—figure supplement 1
Accuracy and confidence of the computer generated partners (CGPs).

Confidence is plotted in blue and accuracy is plotted in red. (a) Study 1 – HAHC: high accuracy and high confidence. HALC: high accuracy and low confidence. LAHC: low accuracy and high confidence. …

Figure 1—figure supplement 2
Statistical analysis of the confidence matching effect.

(a) Top: Permutation test. The empirically observed difference in mean confidence (red line) is significantly different from the distribution of the expected mean (black curve and dotted line) under …

Figure 1—figure supplement 3
Examination of the hypothesis that the partner’s confidence at trial t modulates the participant behavior at trial t+1.

Probability correct: First row, confidence: Second row and reaction time (RT): last row. Study 1: first column. Study 2: second column. We used a generalized linear mixed model (GLMM) similar to …

Figure 1—figure supplement 4
Summary of debriefing results of the second study.

(a) Participants felt that their partner was more confident when facing with a high confidence agent (HCA). This means our manipulation indeed worked. (b) Similar to (a) but for accuracy. Here, …

Figure 2 with 2 supplements
Pupil size during inter-trial interval (ITI) under pairing conditions in the social context when participant was paired with a high (HCA) or low confidence (LCA) agent.

Normalized pupil diameter aligned to start of ITI period (t=0). Vertical dashed lines show average ITI duration. The shaded areas are one standard deviation of ITI period in each condition. Inset …

Figure 2—figure supplement 1
Pupil size correlates with participant’s own confidence in the isolated condition.

(a) In study 1, when confidence was lowest (i.e. rated 1) pupil size was larger (orange curve) than highest confidence (rated 6, magenta curve). The shaded area indicates the average inter-trial …

Figure 2—figure supplement 2
Time series analysis of pupil size during inter-trial interval.

For each study, we employed a generalized linear mixed model (GLMM) Pupil(t)=b0+b1*socialCondition where socialCondition is high confidence agent (HCA) = 1 or low confidence agent (LCA) = 2. The …

Figure 3 with 8 supplements
Neural attractor model.

(a) Left: A common top-down (Wx) current drives both populations, each selective for a different choice alternative. Right: A schematic illustration of the impact of a positive top-down drive on …

Figure 3—figure supplement 1
Confidence matching without removing the correlation with the shared stimulus coherence.
Figure 3—figure supplement 2
The effect of top-down current on the attractor network.

The results of model simulations with a specific value of top-down current (Wx = 0.003). This plot shows the average accumulated evidence of the model in 1000 repetitions with (solid lines) and …

Figure 3—figure supplement 3
Model performance regarding different confidence representations.

(a) Confidence representation based on Equation 15. (b) Same as (a) but here confidence is calculated as the absolute difference of the winner and loser signal but only at the end of the stimulus …

Figure 3—figure supplement 4
Model comparison.

We fit each model to the data from high confidence agent (HCA) and low confidence agent (LCA) blocks of each subject (N=15 (subjects) * 2 (blocks)=30). (a) Pie chart indicates the distribution of …

Figure 3—figure supplement 5
Model vs data.

(a–c) Correspondence between behavioral data (black circles, 40 trials per coherence level in the isolated session) and the model fits (red curves, simulation with 1000 trials per coherence level) …

Figure 3—figure supplement 6
The speed of confidence matching.

(a) Empirical data depicting the time course of confidence matching. The difference between the subject’s confidence and that of the agent (y-axis) are averaged within a three-trial time window and …

Figure 3—figure supplement 7
Model falsification.

(a) Top-down model. We simulated two versions of the model (2000 trials per coherence level) in which only top-down current was different between conditions (TD in high confidence agent [HCA] is …

Figure 3—figure supplement 8
Model predictions for confidence matching are not sensitive to linearity assumptions.

(a) Replication of Figure 3b of the main text in which two models are coupled linearly (see Equation 18) and show confidence matching. (b) Same as (a) but with quadratic coupling. In order to show …

Figure 4 with 5 supplements
Coupling of neural evidence accumulation to social exchange of information.

(a) Centroparietal positivity (CPP) component in the isolated condition: event-related potentials are time-locked to stimulus onset, binned for high and low levels of coherency (for study 1, low: …

Figure 4—figure supplement 1
Electrode placement in each study.
Figure 4—figure supplement 2
Relation of EEG signals from centropartial area of the brain to coherence levels and social conditions.

Top-left, ramping activities of the signals (calculated by a linear regression of signals amplitudes and the time windows of 0–500 ms) is modulated by coherence levels (generalized linear mixed …

Figure 4—figure supplement 3
Simulated slope of the accumulator activity in our computational model in low confidence agent (LCA) and high confidence agent (HCA) conditions.

(a) Slope of the winning accumulator (time window: 0–500 ms; shaded area, insets) at each coherence level for LCA and HCA condition. (b) Same as panel (a) but here for the difference in accumulator …

Figure 4—figure supplement 4
Response-locked EEG signal separated for high vs low coherence levels.

(a) As expected from previous studies (Kelly and O’Connell, 2013; Loughnane et al., 2018; O’Connell et al., 2018; Vafaei Shooshtari et al., 2019) centropareital positivity (CPP) signals to high vs …

Figure 4—figure supplement 5
Power calculation (Monte Carlo simulation) for EEG slope effect (Figure 4 in the main manuscript).

Our power calculator suggests we need 17 participants. EEG slope effect was the only effect that was not statistically significant in the first study.

Author response image 1
Calibration plot for the experimental setup.

Average pupil size (arbitrary units from eyelink device) is plotted against display luminance. The plot is obtained by presenting the participant with uniform full screen displays with 10 different …

Tables

Table 1
Details of statistical results in behavioral data (Figure 1).
ResponseRegressorsEstimateSECIt-Statp-ValueTotal number
Study 1Accuracy
(HC vs LC)
Coherency0.0070.0006[0.006 0.008]11.57<0.0019600
Condition–0.0020.021[–0.045 0.04]–0.10.929600
Confidence
(HC vs LC)
Coherency0.04750.0008[0.046 0.049]56.5<0.0019600
Condition1.3610.03[1.31 1.42]46.4<0.0019600
RT
(HC vs LC)
Coherency–0.0050.0001[–0.005 –0.004]–44.4<0.0019600
Condition0.0290.004[–0.035 –0.021]7.85<0.0019600
Study 2Accuracy
(HC vs LC)
Coherency0.02090.0016[0.017 0.024]13.23<0.0016000
Condition–0.00920.0296[–0.067 0.049]–0.310.766000
Confidence
(HC vs LC)
Coherency0.10110.1011[0.097 0.106]47.47<0.0016000
Condition0.4960.037[0.42 0.56]13.32<0.0016000
RT
(HC vs LC)
Coherency–0.0090.0003[–0.01 –0.008]–26.22<0.0016000
Condition0.03630.006[0.024 0.048]6.12<0.0016000
Table 2
Details of statistical results in pupil data (Figure 2).
ResponseRegressorsEstimateSECIt-Statp-ValueTotal number
Study 1PupilCondition–0.0380.011[–0.06 –0.01]–3.30<0.0018390
Study 2PupilCondition–0.0660.015[–0.09 –0.04]–4.37<0.0015842
Table 3
Details of statistical results in EEG data (Figure 4).
ResponseRegressorsEstimateSECIt-Statp-ValueTotal number
Study 1EEG slopeCoherency0.620.065[0.49. 074]9.64<0.0016492
Condition0.20.14[-0.07 0.49]1.420.156492
Study 2EEG slopeCoherency0.80.29[0.24 1.37]2.8<0.015367
Condition1.520.63[0.27 2.77]2.390.0175367
Table 4
Details of statistical results in EEG data (Figure 4—figure supplement 2 top row).
ResponseRegressorsEstimateSECIt-Statp-ValueTotal number
Study 1EEG slopeCoherency0.020.005[0.01 0.03]4.48<0.0011523
Study 2EEG slopeCoherency0.060.02[0.01 0.11]2.54<0.012822
Table 5
Details of statistical results for the impact of previous trial (Figure 1—figure supplement 3).
ResponseRegressorsEstimateSECIt-Statp-ValueTotal number
Study 1Accuracy
(HC vs LC)
Coherency0.0070.0006[0.006 0.008]11.58<0.0019600
Conf (t–1)–0.00170.005[–0.01 0.01]–0.280.779600
Confidence
(HC vs LC)
Coherency0.0470.001[0.045, 0.049]54.7<0.0019600
Conf (t–1)0.320.008[0.3 0.33]38.31<0.0019600
RT
(HC vs LC)
Coherency–0.0050.0001[–0.0048 0.0044]–44.36<0.0019600
Conf (t–1)–0.00550.001[–0.007 –0.003]–5.44<0.0019600
Study 2Accuracy
(HC vs LC)
Coherency0.020.002[0.02 0.024]13.23<0.0016000
Conf (t–1)0.0030.008[–0.012 0.018]0.370.76000
Confidence
(HC vs LC)
Coherency0.10.002[0.097 0.0106]47.2<0.0016000
Conf (t–1)0.090.01[0.07 0.11]8.6<0.0016000
RT
(HC vs LC)
Coherency–0.0090.0003[–0.001 –0.008]–26.2<0.0016000
Condition0.0050.001[0.001 0.008]2.98<0.016000
Table 6
The rate of trial rejection of eye tracking (only data of social) and EEG data (visual inspection) per participant.
ParticipantsEye tracking rejection % (social)EEG trial rejection % (visual)
Study 1 (Discovery)112.254.6
212.8731.1
30.522.1
4414.8
51.3734.4
604.6
77.758.8
80.3724.4
96.377.6
10046
110.12NA
12NANA
Study 2 (Replication)104
21.251
35.758.5
40.53
5116
61.52.5
700.5
81.59
902
1014
1117.5
120.50
130.7510.5
142.512
1514.754.5
Table 7
Generalized linear mixed model (GLMM) including interaction terms (p-values are reported).
ResponseCoherenceCondition (LC vs HC)Condition* coherence
Study 1Accuracyp<0.001p=0.92p=0.96
Confidencep<0.001p<0.001p<0.001
RTp<0.001p<0.001p<0.05
Pupilp=0.43p=0.20p=0.31
EEG slopep<0.01p=0.15p=0.91
Study 2Accuracyp<0.001p=0.75p=0.87
Confidencep<0.001p<0.001p<0.001
RTp<0.001p<0.001p=0.34
Pupilp=0.35p=0.06p=0.17
EEG slopep=0.62p<0.05p=0.68
Table 8
Attractor model’s parameters.
ParameterParameter valueReference, remarks
JN,ii0.3157 nACalibrated based on pool of isolated data, also fitted on individual subjects’ data
JN,ij0.0646 nACalibrated based on pool of isolated data, also fitted on individual subjects’ data
µ045.8 HzCalibrated based on pool of isolated data, also fitted on individual subjects’ data
NDT0.27 sCalibrated based on pool of isolated data, also fitted on individual subjects’ data
Bound0.32 nACalibrated based on pool of isolated data, also fitted on individual subjects’ data
a (Equation 15)–0.99Calibrated based on pool of isolated data, also fitted on individual subjects’ data
b0 (Equation 15)1.32Calibrated based on pool of isolated data, also fitted on individual subjects’ data
b1 (Equation 15)–0.165Calibrated based on pool of isolated data, also fitted on individual subjects’ data
k (Equation 15)5.9Calibrated based on pool of isolated data, also fitted on individual subjects’ data
I00.3255 nAFrom Wang, 2002; Wong and Wang, 2006
JA.ext0.00022 nA Hz–1From Wang, 2002; Wong and Wang, 2006
τs0.1 sFrom Wang, 2002; Wong and Wang, 2006
dt0.0005 sFrom Wang, 2002; Wong and Wang, 2006
a (Equation 13)270 (V nC)–1From Wang, 2002; Wong and Wang, 2006
b (Equation 13)108 HzFrom Wang, 2002; Wong and Wang, 2006
d (Equation 13)0.154 sFrom Wang, 2002; Wong and Wang, 2006
γ0.641From Wang, 2002; Wong and Wang, 2006
Noise_std0.025From Wang, 2002; Wong and Wang, 2006
I_noise0.02From Wang, 2002; Wong and Wang, 2006

Additional files

Download links