Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2024 Sep 11;44(37):e2236232024.
doi: 10.1523/JNEUROSCI.2236-23.2024.

Decision-Making with Predictions of Others' Likely and Unlikely Choices in the Human Brain

Ning Ma  1   2 , Norihiro Harasawa  1 , Kenichi Ueno  3 , Kang Cheng  3   4 , Hiroyuki Nakahara  5
Affiliations

Decision-Making with Predictions of Others' Likely and Unlikely Choices in the Human Brain

Ning Ma et al. J Neurosci. .

Abstract

For better decisions in social interactions, humans often must understand the thinking of others and predict their actions. Since such predictions are uncertain, multiple predictions may be necessary for better decision-making. However, the neural processes and computations underlying such social decision-making remain unclear. We investigated this issue by developing a behavioral paradigm and performing functional magnetic resonance imaging and computational modeling. In our task, female and male participants were required to predict others' choices in order to make their own value-based decisions, as the outcome depended on others' choices. Results showed, to make choices, the participants mostly relied on a value difference (primary) generated from the case where others would make a likely choice, but sometimes they additionally used another value difference (secondary) from the opposite case where others make an unlikely choice. We found that the activations in the posterior cingulate cortex (PCC) correlated with the primary difference while the activations in the right dorsolateral prefrontal cortex (rdlPFC) correlated with the secondary difference. Analysis of neural coupling and temporal dynamics suggested a three-step processing network, beginning with the left amygdala signals for predictions of others' choices. Modulated by these signals, the PCC and rdlPFC reflect the respective value differences for self-decisions. Finally, the medial prefrontal cortex integrated these decision signals for a final decision. Our findings elucidate the neural process of constructing value-based decisions by predicting others and illuminate their key variables with social modulations, providing insight into the differential functional roles of these brain regions in this process.

Keywords: computational modeling; dorsolateral prefrontal cortex; fMRI; posterior cingulate cortex; predicting other individuals; social decision-making.

PubMed Disclaimer

Conflict of interest statement

The authors declare no competing financial interests.

Figures

Figure 1.
Figure 1.
Experimental task. a, The experimental task consisted of three types of trials, one main trial (cued by a rectangle with circle inside at the beginning) and two control trials (control-choice and control-prediction trials, cued by a rectangle and a circle, respectively). Subjects were instructed to choose between two options, each of which was associated probabilistically with rewards, in the main and control-choice trials to maximize their gains and in the control-prediction trials to predict another person's choice. Each option was provided with one orientation bar and two numbers in the center at the top and bottom. Reward probability was indicated by the orientation of the bar (see panel b) and reward magnitudes were indicated by the number. In the main trial, which number at the top or bottom indicated the magnitude in a trial was determined by the other's choice (see panel c for details). In the control-choice trial, the number with the same color as the rectangular border of the cue would be rewarded. In this example, the cue is red, so the actual reward magnitudes were 20 and 40 for the left and right options, respectively. In the control-prediction trial, the reward magnitudes for others were always the top numbers regardless of the color. In this example, this is 20 for left and 60 for right. At the end of each trial, the outcome was shown. In this example, the subject chose the right option and was finally rewarded in the control-choice trial (indicated by an orange circle, 40 points) and in the control-prediction trial (correctly predicted the other's choice indicated as white circle on the right option and received a constant 30 points) but was not rewarded in the main trial (indicated by a black cross). All three trial types were interleaved in a block of trials. b, Reward probability was sampled from five different probabilities, corresponding to different orientations of a bar. Here, in this orientation-probability map, a horizontal bar indicated the minimal probability of 20% while a vertical bar indicated the maximal probability of 80%. Bars with the three other possible orientations in the middle indicated probabilities of 35, 50, and 65%, respectively. The white numbers for reward probabilities here were not shown in the trials in the experiments, while the tick marks around the circle were always shown to help subjects to distinguish the different orientations. Different probability maps were set for self and others for a given participant, so that the same orientation bar can indicate different reward probabilities for the self and others. For details, see Materials and Methods, Setting of the bar orientation-reward probability map. c, In the main trials, others’ reward probability was determined from the orientation-probability map for others, and others’ reward magnitudes were always shown by the numbers on top of the two options. Here, from the options in panel a, the numbers were 20 for left and 60 for right. For self-options, for the option chosen by others, the reward magnitude used was the bottom number, whereas, for the option unchosen by others, the reward magnitude was the top number. Here, if the left option is chosen by others, the self-reward magnitudes are 40 for left and 60 for right. Otherwise, if the right option is chosen by others, the self-reward magnitudes are 20 for left and 40 for right. Therefore, to make better decisions for the self, subjects should predict others’ choices and figure out the reward magnitude for themselves.
Figure 2.
Figure 2.
Behavioral results. a, Control-choice and control-prediction trials. Subjects’ choices were modulated by the corresponding value differences. Left and middle panels, Their behaviors were plotted against the respective value difference, shown as the mean with standard error across subjects. Right panel, The effect size (d’) for detecting the higher values in the left options was presented according to the signal detection theory, wherein the bar and error bar indicate the means and standard errors, respectively, with dots indicating the d’ of each subject (and similarly hereafter). The respective value difference showed significantly positive effects (d’ > 0, all p values <0.001) on their choice behavior in control-choice and control-prediction trials, with no significant difference in between (paired t test on d’; t(47) = −0.749; p = 0.457). b, Main trials. Subjects made their choices, relying on the use of value differences that were appropriate based on their predictions of others’ choices. Subjects’ choices followed two value differences for one's own rewards, as in each panel, plotted against ΔSV(O = L; blue) and ΔSV(O= R; red), articulated by the value difference for others, as shown across three panels, with three trial groups of almost equal sizes sorted by the magnitudes of ΔOV. The behavior was differentially well modulated by ΔSV(O = L) and ΔSV(O = R) when [ΔOV> 0] and [ΔOV < 0], respectively. All of the values were estimated by the best models from the control trials, and the differences were defined by the left minus the right. c, Top, Statistics for panel b; effect sizes (d’) of the six plots in panel b are shown; *, significant difference between the two effects ([ΔOV < 0], t(47) = −6.954, p < 0.001; [ΔOV > 0], t(47) = 8.525, p < 0.001). Bottom, The differential effects of the two value differences [ΔSV(O = L) and ΔSV(O = R)] were manifested through the difference in reward magnitudes. Effect sizes (d’) of the difference in reward probability (Δp) and difference in reward magnitudes [Δm(O = L) and Δm(O = R)]; *, significant difference between the two effects (for Δm, [ΔOV < 0], t(47) = −4.983, p < 0.001; [ΔOV > 0], t(47) = 6.067, p < 0.001). d, Subject's choices were influenced by both types of the others’ value difference, especially when the prediction was difficult. Effect sizes (d’) of primary and secondary value differences (pΔSV and sΔSV, top panel) and primary and secondary reward magnitude differences (Δmp and Δms, bottom panel) are shown for three subgroups of trials (all trials and the top 50% and top 10% of the most “difficult-to-predict” trials). *, significant with one-way repeated-measures ANOVA (for sΔSV: F(2,94) = 9.61, p < 0.001; for Δms: F(2,94) = 5.351, p = 0.006); n.s., no significant difference. e, The fit to the choice behavior in the main trials is shown (AIC values, each of which indicates the averaged value across the trials for an individual participant) for the three categories of models based on Assumptions I, II, or III (abbreviated AI, AII, and AIII). AI, only primary values contributed to behavior; AII, both primary and secondary values contributed; AIII, without prediction of others’ choices. Violin plots indicate the distributions of all subjects (N = 48), with black and red lines representing the mean and median. To indicate each subject in the figure, a symbol was determined based on the best of the three models, namely, red “x”, blue “+”, or black “o” if the AI, AII, or AIII model was the best, respectively. The same symbol was used for each subject across the three models, connected by the line with the same color. Group-wisely, the AI or AII model was significantly better for fitting behavior than the AIII model (paired t test of AIC values: I vs III, t(47) = −7.461, p < 0.001; II vs III, t(47) = −8.688, p < 0.001) while the AII model was better for fitting behavior than the AI model (I vs II, t(47) = 4.446; p < 0.001). These results were from the data of all experiments (fMRI + behavioral), and the same results were also found (ps < 0.001) by using only the data from the fMRI experiments. f, Left panel, Numbers of the subject best fit by each model. Right panel, Percent changes of the difference in fit between the AI and AII models, sorted across subjects, not including subjects who were best fit by an AIII model. We observed a large variation but relatively gradual changes across the subjects, suggesting the behaviors were on a continuum rather than in distinct categories.
Figure 3.
Figure 3.
BOLD signals by GLM analysis. a, Decision signals in the main and control-choice trials. Activations of decision values (ΔDV, the value of chosen options minus that of the unchosen ones, extracted by GLM2 for all Group I–III subjects) were found in the medial prefrontal cortex (mPFC). They survived at the FWE-corrected p < 0.05 level, and the same criteria are used in the following figures (or stated otherwise). For display purpose, here and hereafter, individual voxels in the activation map were first subjected to thresholding at p < 0.005 and then blurred at the resolution of anatomic imaging. b, Prediction signals in the main and control-prediction trials. Activations of others’ choice probability for the chosen options (extracted by GLM3 using Group I and II subjects) were found in the left amygdala. Although the activation center in main trials is not visually within the left amygdala, the activation volume overlaps the left amygdala by eight voxels according to Talairach coordinates. c, Signals for differences in reward magnitudes in the main trial. With GLM1 (using both Group I and II subjects for Δmp and using only Group II subjects for Δms), the posterior cingulate cortex (PCC) was activated by Δmp, the primary reward magnitude difference (chosen vs unchosen) corresponding to the case in which others chose an option likely to be chosen. The right dorsolateral prefrontal cortex (rdlPFC) was activated by Δms, the secondary reward magnitude difference corresponding to the other case in which others chose an option unlikely to be chosen.
Figure 4.
Figure 4.
Temporal dynamics of brain signals. a–d, Time course of brain signals in the main trials. For each of the two groups of subjects (Group I, blue; Group II, red), the mean and standard error were calculated and are shown as solid lines with shadings. Colored asterisks on the top parts of each panel indicate a significant positive difference against zero (one-sample t test; p < 0.05) for each group, while black asterisks indicate the differences between the two groups (independent samples t test; p < 0.05). The origin on the horizontal axis is set as the onset of the options. e, Significantly positive time periods are summarized for each ROI of each group. The colored asterisks in panels a–d are shown here as small black dots. Colored diamonds indicate the average of all significantly positive times for each ROI.
Figure 5.
Figure 5.
Functional connectivity by PPI analysis. Activation maps by voxel-wise PPI are shown here for the functional connectivities of the left amygdala with the PCC and the rdlPFC and of the PCC and rdlPFC with the mPFC, overlapped on ROIs generated from GLM activations (Fig. 3), which are shown here as yellow backgrounds. For display purposes, the activation identified by voxel-wise PPI is shown by an uncorrected p < 0.005, and the significance of the activation is tested using small-volume correction (SVC) within the target ROIs at the p < 0.05 level (FWE corrected). Bar plots show the ROI PPI results in the main and control trials, wherein * indicates significant differences between the main and control-prediction or control-choice trials. Top left and bottom left panels are the functional connectivities of the left amygdala with the PCC (t(47) = 3.967; p < 0.001; main vs control-prediction trial) and with the rdlPFC (t(47) = −6.968; p < 0.001; main vs control-choice trial), respectively. Top right and bottom right panels are the functional connectivity of the PCC with the mPFC (t(47) = 3.955; p < 0.001; main vs control-choice trial) and of the rdlPFC with the mPFC (t(47) = 2.699; p = 0.010; main vs control-choice trial), respectively.
Figure 6.
Figure 6.
Brain signals across behavioral variations in individuals. a, BOLD signals in the rdlPFC for the secondary reward magnitude difference (Δms) are significantly positively correlated with individuals’ behavioral variations assessed by fitting variations between the AI and AII models (across all subjects from Group I and II; ρ = 0.337; N = 43; p = 0.028; by Spearman correlation). Behavioral variations were quantified by the AIC value difference between the AI and AII models and are rank-ordered in the figure. The rdlPFC signals were extracted from the 6 to 8 s duration from the onset of options, in line with our GLM findings. The same analysis of the signals of the 4–6 s duration was also significant (ρ = 0.397; N = 43; p = 0.009), thus confirming the current result. b, Because the rdlPFC activations contributed to the use of the Δms, we further investigated the relation between behavior and functional connectivities of the rdlPFC in Group II subjects. The PPI effect size of the left amygdala with the rdlPFC is significantly positively correlated with the individuals’ behavioral variations within Group II subjects (ρ = 0.400; N = 29; p = 0.033).
See this image and copyright information in PMC

Similar articles

See all similar articles

References

    1. Adolphs R (2002) Neural systems for recognizing emotion. Curr Opin Neurobiol 12:169–177. 10.1016/S0959-4388(02)00301-X - DOI - PubMed
    1. Adolphs R, Gosselin F, Buchanan TW, Tranel D, Schyns P, Damasio AR (2005) A mechanism for impaired fear recognition after amygdala damage. Nature 433:68–72. 10.1038/nature03086 - DOI - PubMed
    1. Alcala-Lopez D, et al. (2018) Computing the social brain connectome across systems and states. Cereb Cortex 28:2207–2232. 10.1093/cercor/bhx121 - DOI - PubMed
    1. Amodio DM, Frith CD (2006) Meeting of minds: the medial frontal cortex and social cognition. Nat Rev Neurosci 7:268–277. 10.1038/nrn1884 - DOI - PubMed
    1. Baron SG, Gobbini MI, Engell AD, Todorov A (2011) Amygdala and dorsomedial prefrontal cortex responses to appearance-based and behavior-based person impressions. Soc Cogn Affect Neurosci 6:572–581. 10.1093/scan/nsq086 - DOI - PMC - PubMed

LinkOut - more resources