Social Decision Preferences for Close Others Are Embedded in Neural and Linguistic Representations

Participants

A total of 63 (mean age, 18.81 years; SD, 1.05 years; 73% female) individuals participated in the current study. Participants were recruited from the University of California, Los Angeles (UCLA) student body by tapping existing internal recruitment pools, classroom advertisements, and an email from the registrar's office. Our a priori sample size goal was to recruit and run as many participants as possible during the winter and spring academic quarters (January to June 2023). In order to be eligible for the study, participants must have been between 18 and 22 years at the time of enrollment and exhibit no MRI contraindications (e.g., claustrophobia, non-MR safe metallic implants, etc.). All participants were fluent in English. Four participants did not have imaging data available (three due to software issues preventing data collection/reconstruction and one due to participant failure to fully disclose permanent jewelry during safety screening), rendering a sample size of N = 59 for imaging analyses. All participants provided consent in accordance with the policies of the UCLA Institutional Review Board. Participants were compensated $40 (USD) upon completion of the study. All data and materials are publicly available on the Open Science Framework (OSF; osf.io/fzds6).

Demographically, the full sample was comprised of 46 individuals assigned the female sex at birth (73.0%) whereas the rest were assigned male. In terms of gender, 43 individuals identified as women (68.3%), 17 as men (27.0%), 2 as nonbinary (3.2%), and 1 (1.6%) declined to respond. Ethnically, 12 identified as Hispanic/Latinx (19.0%). Racially, 24 identified as Asian/American (38.1%), 23 identified as Caucasian (36.5%), 7 identified as mixed race/multiethnic (11.1%), 5 identified as another category (“Other,” 7.9%), 4 declined to respond (6.3%), 0 identified as Black (0%), and 0 identified as Native Hawaiian/Pacific Islander and American Indian/Alaska Native (0%). In terms of sexual orientation, 46 individuals identified as straight/heterosexual (73.0%), 8 identified as bisexual (12.7%), 4 identified as homosexual (gay/lesbian, 6.3%), 3 identified as another category (“Other,” 4.8%), and 2 identified as asexual (3.2%). Finally, 57 participants reported relying on their parents for financial assistance (90.5%). The self-reported median household income among participants was $100,000 (range, $18,000–$3,000,000).

Protocol overview

Participants completed the study in a 1 h session at UCLA's Staglin One Mind Center for Cognitive Neuroscience. Participants were greeted by an experimenter upon arrival, consented, nominated a parent and close friend, received training for the two fMRI tasks, underwent task-based fMRI scanning, and then completed a postscan survey. Major elements of this procedure are described in greater detail below.

Parent–close friend nomination

Prior to receiving training for the fMRI tasks, participants were informed the purpose of the study was to understand how the brain encodes information about close others and how it uses such information when making choices about others. Participants were then asked to nominate a parent and their closest friend. The closest friend was not allowed to be a current romantic partner or a family member. For the parent, the majority of participants chose a mother (84%) whereas the remainder chose a father (mean age of nominated parent, 51.4 years; SD, 5.6). The majority (69%) of participants nominated a close friend assigned the female sex at birth (mean age of nominated friend, 19.1 years; SD, 1.2 years).

fMRI acquisition parameters

fMRI data were collected using a research-dedicated 3 T Siemens Prisma scanner and a 32-channel head coil. Functional scan sequences were designed to be consistent with the sequences used by the Human Connectome Project (humanconnectome.org), which uses the University of Minnesota's Center for Magnetic Resonance Research (CMRR) multiband accelerated EPI pulse sequences (github.com/CMRR-C2P/MB). A high-resolution T1 magnetization-prepared rapid-acquisition gradient echo structural image was acquired for registration purposes (TR, 1,900 ms; TE, 2.48 ms; flip angle, 9°; FOV, 256 mm²; 1 mm³ isotropic voxels; 208 slices; A>>P phase encoding). Functional runs were comprised of T2*-weighted multiband echoplanar images (TR, 1,240 ms; TE, 36 ms; flip angle, 78°; FOV, 208 mm²; 2 mm³ isotropic voxels; 60 slices; A>>P phase encoding; multiband acceleration factor, 4). Single-band reference images were acquired for each functional run.

Behaviors and preferences judgment task

Mental representations of the participant (self), close friend, and parent were elicited during fMRI scanning with a computerized behaviors and preferences judgment task (Tamir and Mitchell, 2013; Thornton and Mitchell, 2018). The task entails rating how well a series of statements encompassing a variety of behaviors and preferences describe a given target. On each trial, a label for a given target was displayed at top of the screen ('SELF,” “PARENT,” or “FRIEND”), a preference was displayed in the center of the screen, and a four-point Likert scale was displayed at the bottom (1, “Strongly disagree”; 4, “Strongly agree”). Examples of preferences include “really enjoys living in big cities,” “thinks modern art is uninteresting,” and “feels comfortable talking about personal issues.” Participants used a button box to indicate their response. Each trial was presented for 4,500 ms, and was followed by a jittered fixation screen consisting of a black “+” in the center of the screen (mean = ∼2,000 ms; possible range = [200–10,165 ms]; distributed exponentially). Trials were “chunked” by a target such that participants would rate at least seven consecutive trials about a given target before moving on to another target to help reduce the cognitive load of rapidly switching between targets. To help ensure participants did not miss a transition between chunks of targets, the fixation cue at the center of the screen was displayed in red prior to the start of a new chunk.

Participants completed 225 trials total (75/target), spread across four runs. The number of trials per target varied across runs. Importantly, the exact same set of stimulus items (preferences) was used for each target (self, parent, friend), helping ensure that representations of oneself, a parent, and friend were comparably elicited and that different patterns of effects between self–parent and self–friend similarities would not be driven by stimulus differences. Each run typically lasted between 345 and 375 s, given the random jitter. Preferences for this task were drawn from an open, online pool used in prior studies (https://jasonmitchell.fas.harvard.edu/links/; accessible on our OSF repository at osf.io/sfc6b). Using methods consistent with prior studies (Tamir et al., 2016), we narrowed the initial 800+ item pool into a final set of 75 items based on how strongly they were associated with Big Five personality traits (John and Srivastava, 1999; De Raad, 2000). We chose to use the Big Five as an organizing framework to structure our measurement of participant representations because personality traits are a key component of individual identities, are known to be broadly generalizable, and encompass other key types of information (e.g., they imply information about mental state–scenario associations). The optimization procedure involved applying a genetic algorithm to stimulus item ratings solicited in an independent, crowd-sourced sample. Further stimulus optimization details are hosted on our OSF page (osf.io/zcu7n). We checked whether reaction time data from the task differed between condition pairs of interest (self, parent; self, friend) while accounting for random effects of subject and stimulus by using hierarchical modeling. No such differences emerged. The task was programmed in jsPsych (de Leeuw et al., 2023).

Social decision-making task

We used a pair of computerized delay discounting tasks to assess social decision preferences between parents and close friends. Our rationale for choosing a discounting task is threefold. First, discounting decisions are both pervasive in everyday life and are thought to be important for shaping life adjustment outcomes (Golsteyn et al., 2014; Lebeau et al., 2016). Second, computerized discounting tasks deployed in research settings are flexible in their configuration, allowing researchers to study social decision behavior in the face of different reward outcomes (Seaman et al., 2016). Last, we have previously shown that discounting tasks with social choice manipulations reliably tap decision preferences between specific social partners and are related to socioemotional facets of interpersonal relationships such as quality and subjective relationship value (Guassi Moreira and Parkinson, 2024); we also previously observed that choice preferences from similar tasks tend to be correlated (Guassi Moreira et al., 2020, 2021).

Participants completed two separate runs of a delay discounting task. On each trial, participants were presented with two hypothetical scenarios that pitted outcomes for their nominated parent and friend against each other. One scenario involved a relatively immediate, smaller reward, and the other involved a relatively delayed, larger reward. The delays could take the value of zero (“Today”), 2 weeks, 4 weeks, and 6 weeks. Values of zero and 6 weeks were never presented in the delayed or immediate scenarios, respectively. The two runs differed in the type of rewards offered to participants. One of the runs offered participants hypothetical monetary rewards (in USD) to be earned on behalf of either nominee (e.g., $16 for a given target); the other run was comprised of social rewards, offering participants hypothetical time spent with either target (e.g., 16 min of time spent with a given close other). Participants were instructed to think of the latter condition as “additional” time they could spend with their parent or friend, combined with the time they already spend with either close other.

Numeric outcome values for each type of reward ranged from 2 to 30. There were 49 unique combinations of numeric values and time pairings (e.g., $2 now vs $18 2 weeks from now) for each condition, resulting in 294 unique possible pairings (Guassi Moreira et al., 2021). In the interest of reducing in-scan task demands for participants, each run was comprised of 30 pairings randomly selected from the total 294. It was heavily stressed that participants were to complete the task as if rewards were real, even though they were hypothetical. Participants had 4,500 ms to indicate their decision via button box. The trial terminated once they completed their response. Each trial was interspersed with a fixation stimulus “+” and a random jitter (similar parameters as the previous task). Participants completed two runs of the task (one monetary, one social). Each run lasted ∼90–135 s, with variability owing to the self-paced nature of the task. The visual configuration of task stimuli was generally consistent with prior implementations (Seaman et al., 2016; Guassi Moreira et al., 2021). Both runs were programmed in PsychoPy (Peirce et al., 2019). This task was always presented to participants after the behaviors and preferences judgment task. Outcome types for this task were counterbalanced between participants.

Postscan survey

Participants completed a postscan survey immediately following the scan. During the survey, they provided basic demographic information (sex and age) about their nominated parent and friend, were asked to list ∼10 adjectives or short phrases describing each close other, and answered four written prompts asking them to recall specific memories with each close other. The prompts asked participants to write about their (1) favorite memory with each close other, (2) a time when their close other made them feel supported, (3) a time when their close other let them down, and (4) and a time when they let their close other down. Participants were asked to provide 3–5 sentences per each memory with as much detail as possible. Text about parents and friends were collected in separate blocks. Block orders were randomized between participants. The survey also contained measures of their relationship quality with both close others (see below, Study 2), questions about how often they see each close other (actual and ideal), a measure of social loss aversion toward each close other (i.e., how upset one would be if they could no longer spend time with a given social partner; Guassi Moreira and Parkinson, 2024), and two one-shot social decision-making items (a dictator game and a forced-choice question about whom they would rather spend time with). More information on these latter measures is fully disclosed on our OSF page (osf.io/zcu7n). Participants finally finished the survey by answering demographic items about themselves.

MRI data preprocessing

fMRIPrep (version 22.0.2)—an open source, freely available tool based on functional neuroimaging software Nipype (Gorgolewski et al., 2011)—was used to perform preprocessing MRI preprocessing. The T1-weighted (T1w) image was corrected for intensity nonuniformity with N4BiasFieldCorrection (Tustison et al., 2010), distributed with ANTs 2.3.3 (Avants et al., 2008; RRID:SCR_004757), and used as T1w reference throughout the workflow. The T1w reference was then skull-stripped with a Nipype implementation of the antsBrainExtraction.sh workflow (from ANTs), using OASIS30ANTs as target template. Brain tissue segmentation of cerebrospinal fluid (CSF), white matter (WM), and gray matter (GM) was performed on the brain-extracted T1w using fast (FSL 6.0.5.1; Zhang et al., 2001). Brain surfaces were reconstructed using recon-all (FreeSurfer 7.2.0; Dale et al., 1999), and the brain mask estimated previously was refined with a custom variation of the method to reconcile ANT-derived and FreeSurfer-derived segmentations of the cortical GM of Mindboggle (Klein et al., 2017). Volume-based spatial normalization to one standard space (MNI152NLin2009cAsym) was performed through nonlinear registration with antsRegistration (ANTs 2.3.3), using brain-extracted versions of both the T1w reference and the T1w template. The following template was selected for spatial normalization: “ICBM 152 Nonlinear Asymmetrical template version 2009c” (TemplateFlow ID: MNI152NLin2009cAsym; Fonov et al., 2009).

For each of the BOLD runs found per subject, the following preprocessing was performed. First, a single-band reference was used as a reference volume. Head-motion parameters with respect to the BOLD reference (transformation matrices and six corresponding rotation and translation parameters) were estimated before any spatiotemporal filtering using mcflirt (FSL 6.0.5.1; Jenkinson et al., 2002). The BOLD time series were resampled onto their original, native space by applying the transforms to correct for head motion. These resampled BOLD time series will be referred to as “preprocessed BOLD in original space” or just “preprocessed BOLD.” The BOLD reference was then coregistered to the T1w reference using bbregister (FreeSurfer) which implements boundary-based registration (Greve and Fischl, 2009). Coregistration was configured with six degrees of freedom. First, a reference volume and its skull-stripped version were generated using a custom methodology of fMRIPrep. Several confounding time series were calculated based on the “preprocessed BOLD”: framewise displacement (FD), DVARS, and three region-wise global signals. FD was computed using two formulations following Power (absolute sum of relative motions; Power et al., 2014) and Jenkinson (relative root mean square displacement between affines; Jenkinson et al., 2002). FD and DVARS are calculated for each functional run, both using their implementations in Nipype (following the definitions by Power et al., 2014). The three global signals are extracted within the CSF, the WM, and the whole-brain masks. Additionally, a set of physiological regressors were extracted to allow for component-based noise correction (CompCor; Behzadi et al., 2007). Principal components were estimated after high-pass filtering the preprocessed BOLD time series (using a discrete cosine filter with 128 s cutoff) for the two CompCor variants: temporal (tCompCor) and anatomical (aCompCor). tCompCor components were then calculated from the top 2% variable voxels within the brain mask. For aCompCor, three probabilistic masks (CSF, WM, and combined CSF + WM) are generated in anatomical space. The implementation differs from that of Behzadi et al. (2007) in that instead of eroding the masks by 2 pixels on BOLD space, a mask of pixels that likely contain a volume fraction of GM is subtracted from the aCompCor masks. This mask was obtained by dilating a GM mask extracted from the FreeSurfer's aseg segmentation, and it ensures components are not extracted from voxels containing a minimal fraction of GM. Finally, these masks are resampled into BOLD space and binarized by thresholding at 0.99 (as in the original implementation). Components are also calculated separately within the WM and CSF masks. For each CompCor decomposition, the k components with the largest singular values are retained, such that the retained components’ time series are sufficient to explain 50% of variance across the nuisance mask (CSF, WM, combined, or temporal). The remaining components were dropped from consideration. The head-motion estimates calculated in the correction step were also placed within the corresponding confounds file. The confound time series derived from head motion estimates and global signals were expanded with the inclusion of temporal derivatives and quadratic terms for each (Satterthwaite et al., 2013). Frames that exceeded a threshold of 0.5 mm FD or 1.5 standardized DVARS were annotated as motion outliers. Additional nuisance time series were calculated by means of principal components analysis (PCA) of the signal found within a thin band (crown) of voxels around the edge of the brain, as proposed by Patriat et al. (2017). The BOLD time series were resampled into standard space, generating a “preprocessed BOLD run in MNI152NLin2009cAsym space.” First, a reference volume and its skull-stripped version were generated using a custom methodology of fMRIPrep. All resamplings can be performed with a single interpolation step by composing all the pertinent transformations (i.e., head-motion transform matrices, susceptibility distortion correction when available, and coregistrations to anatomical and output spaces). Gridded (volumetric) resamplings were performed using antsApplyTransforms (ANTs), configured with Lanczos interpolation to minimize the smoothing effects of other kernels. Nongridded (surface) resamplings were performed using mri_vol2surf (FreeSurfer).

Multi-voxel pattern estimation

Multi-voxel patternswere estimated by modeling imaging data from the preferences task using a general linear model (GLM). Each run of the task was modeled using a fixed-effect GLM in FSL at the run level. Every GLM consisted of the following explanatory variables: A regressor for each of the self, parent, and friend trials where participants rated the extent to which a given item described the said target. The duration of each regressor corresponded to the length of the trial (4,500 ms). These regressors were convolved with the hemodynamic response function (double gamma). The temporal derivatives of each were added to account for slice timing effects. The following confound regressors were included to statistically control for noise and head motion. Time series from the CSF + WM mask automatically extracted by fMRIPrep was added into the model, in addition to the tCompCor component. The 24 head-motion regressors and individually censored volumes exceeding the FD and DVARS thresholds, both computed by fMRIPrep, were also included. Three linear contrasts were computed: (self-baseline), (parent-baseline), and (friend-baseline).

Representational similarity analysis

We used representational similarity analysis (RSA; Kriegeskorte et al., 2008) to summarize neural representations of others. RSA is useful for comparing representational structure across modalities; in the context of fMRI studies, it is helpful for determining how various forms of psychological or behavioral information are encoded in the brain. Here we use RSA to examine how information about parents and friends is encoded in the brain (specifically self, other overlap), as well as how this information is relevant for social decision-making. The former set of analyses involved conducting various searchlight and ROI analyses; the latter used individual representational dissimilarity matrix (RDM) cells to estimate statistical relationships between individual differences in neural self, other overlap, and social decision preferences. These analyses are described in the following three sections.

A priori ROI analyses

We implemented RSA on region-of-interest (ROI) data to determine whether there was a “bias” in the encoding of parent or friend representations in six brain regions (mPFC, dmPFC, L NAcc, R NAcc, L TPJ, R TPJ). ROIs were defined a priori via meta-analytic maps (peak coordinates) for cortical regions and probabilistic atlases for subcortical regions (NAcc). We constructed RDMs based on between-run distances in multi-voxel patterns of brain activity. Between-run distances were used to avoid the potential confound of temporal “pattern drift,” which has been shown to persist even after voxel-wise detrending has been performed but which can be alleviated by focusing on comparisons between patterns collected in different runs (Alink et al., 2015). Focusing on between-run multi-voxel pattern comparisons is also recommended in designs such as that used in the current study, where particular trials (e.g., those corresponding to a particular target) are “chunked” together (Mumford et al., 2014). Concretely, this meant that multi-voxel patterns for self, parent, and friend were extracted for a given ROI (using t statistic images) and distances were computed between patterns from each run for all possible target pairings. Because we were using between-run distances, this also meant comparing between-run patterns for the same target (e.g., the self-condition for Run 1 was compared with the same condition in Run 2). We ran analyses with both (1 − Pearson's r) and Euclidean distances as the RDM-defining distance metric to check for robustness—analyses were consistent between the two metrics. Euclidean distances are reported in the main document. The mPFC, dmPFC, and TPJ ROIs were defined by locating the respective brain regions on the Neurosynth meta-analytic platform (Yarkoni et al., 2011), noting coordinates for the global maximum (or within each hemisphere for the TPJ), and drawing a 5 mm radius sphere around the coordinates. The NAcc masks were defined using the Harvard–Oxford subcortical atlas as available in FSL. ROI masks are visualized in Extended Data Figure 1-1. We then computed pairwise differences between the (self and parent) and (self and friend) RDM cells for each of the six ROIs. Paired differences were modeled as being drawn from a Gaussian distribution such that Diff_i ∼ N(δ*σ, σ²). The mean was parameterized as δ*σ so that draws from the distribution are in “native” units but summary statistics of paired differences are in Cohen's d metric (i.e., mean/standard deviation). Prior distributions were assigned to both δ [Cauchy (0, 12 )] and σ (Jeffreys prior, proportional to 1σ2 ). The Bayesian modeling software stan (Stan Development Team, 2020) was used to run these analyses via the rstan R package.

Exploratory searchlight analyses

We conducted a set of exploratory searchlight analyses (Kriegeskorte et al., 2006) to identify which brain regions showed the greatest differentiation among the three representations (self, parent, friend) in a data-driven manner. The searchlight was meant to complement the ROI approach by ensuring we did not miss any additional brain regions that differentiated between targets. A 125 mm³ sphere (2 mm radius, excluding center voxel [(2 × 2) + 1]³) was moved throughout each participant's brain, and multi-voxel patterns to each target were compared at each searchlight center. The first comparison we performed was a direct comparison between parent and friend representations. This was accomplished by extracting multi-voxel patterns for the parent and friend t statistics at each center voxel, vectorizing the voxels within the sphere, and taking the (Pearson's) correlation distance (1 − r) between the two patterns. As before, we only used between-run comparisons. We used between-run distances within-target as a baseline. All between-run distances were placed into a vector and standardized. The mean of within-target distances [e.g., (parent Run 1, parent Run 2), (friend Run 2, friend Run 4)] was then subtracted from the mean of (parent, friend) distances. We conducted a second series of analyses comparing representational overlap between the self and each target. Specifically, distances between neural representations of one's self and one's parent were compared with distances between representations of one's self and one's friend in a procedure similar to that described above.

Modeling social decision preferences

We used Bayesian statistics for all regression models involving choice behavior on the social decision-making tasks. Given that there exists notable intra- and inter-individual variability in decision-making behavior, we chose to model the data at the trial level, necessitating the use of a hierarchical logistic regression model. While a hierarchical model is not strictly needed to handle nested data (McNeish et al., 2017), we use it here because it allows us to flexibly model both within- and between-participant sources of variance while regularizing the model coefficients to enhance out-of-sample generalizability. The form of the within-person component of the model is notated below:Logit(Decij)=b0j+b1jConditionij+b2jRewardRatioij. This is the core component of the model because it encodes social decision preferences between parents and friends. Here, the ith decision (Dec, 1, discount; 0, nondiscount) from the jth participant was modeled as a function of a Condition dummy code and a Reward Ratio variable. Condition is coded such that 0 is equal to the friend outcome associated with discounting option, parent outcome associated with nondiscounting (delayed) option, and 1 is equal to the parent outcome associated with discounting option, friend outcome associated with nondiscounting option. The coefficient associated with Condition (b_1j) encodes social decision preferences between parents and friends because it reflects the mean difference in choice tendencies between the two conditions (after statistically adjusting for other terms in the model). A b_1j coefficient > 0 indicates a preference toward parents: there is greater likelihood of choosing to discount when the parent outcome is associated with the smaller immediate reward and a greater likelihood of not discounting when parents are affected by the larger delayed reward. A b_1j coefficient < 0 indicates a friend-oriented preference following the same logic: a reduced likelihood of choosing to discount when the friend outcome is associated with a larger delayed reward and an increased tendency of discounting when associated with the smaller immediate reward. Reward Ratio is the quotient of the nondiscounted outcome over the discounted outcome and is a key covariate for several reasons. Primarily, it serves as a “yardstick” to gauge whether representational information is meaningfully related to decision preferences beyond trial-level features. This is notable because various cognitive heuristics used in decision-making are typically applied over such features; adjusting for this here helps ensure that representational information is still related to choice behavior over and above other cognitive processes. In addition, controlling for Reward Ratio helps to make choices more comparable between trials and provides a “sanity check” for participant engagement (greater reward ratio should generally be associated with a reduced likelihood of discounting regardless of condition). Notably, a post hoc analysis revealed that trial-level task features (e.g., reward ratio, delay ratio, magnitude of rewards) were not significantly oversampled in either choice condition, ruling out the possibility that social decision preferences were not conflated with trial-level choice option features. All three terms in the within-person component of the model were allowed to vary randomly across participants.

Allowing the b_1j coefficient to vary randomly across participants meant that we could model variability in individual decision preferences. We did so across various models by entering different between-participant variables. These between-participant variables are statistically allowed to interact with the Condition dummy code (i.e., a cross-level interaction in a one-level model) but are substantively conceptualized as the association between participant-level individual differences and social choice preferences. The following sets of between-person predictors were tested (each group in a separate model): (1) dissimilarity values for self and parent, as well as self and friend calculated from behavioral ratings, (2–5) neural dissimilarities from each ROI (bilateral NAcc and TPJ ROIs were entered in the same model). Here we notate the between-participant portion of the model with generic neural dissimilarity values as an illustrative example (interested readers can consult our publicly available code for further model details):b0j=γ00+γ01Self,ParentDissimj+γ02Self,FriendDissimj+u0j, b1j=γ10+γ11Self,ParentDissimj+γ12Self,FriendDissimj+u1j, b2j=γ20+u2j. Analyses involving linguistic preference scores (described in the next section) were entered as part of a three-way interaction between neural dissimilarity and Condition in subsequent analyses. All continuous predictors were grand mean centered prior to running the model. All models were implemented using the brms package in the R statistical software library (no thinning, 8 chains, 2,500 samples/chain, 1,250 discarded burn-in samples). A weakly informative standard normal prior was placed on all slope coefficients, and brms default priors were used for all other model parameters.

Because the output of Bayesian analyses is a distribution, we performed statistical inference in a graded manner and thus do not report nor use frequentist p values. Instead, we evaluated the weight of statistical evidence using a combination of the region of practical equivalence (ROPE) method (Kruschke, 2011, 2013) and noninferiority testing (Spiegelhalter et al., 1994; Wiens, 2002). Traditionally, the ROPE method entails specifying a symmetric region in parameter space around the null value, calculating a credible interval (CI), and then evaluating whether the CI intersects at all with ROPE. CIs that do not overlap with ROPE result in a rejection of the null value, CIs that are entirely contained within ROPE result in practical acceptance of the null value, and any other configuration falls in undecided. Originating in clinical trials research, noninferiority testing is similar: if the CI partly overlaps with the upper limit of ROPE but does not pass through below the lower limit (regardless of its relation to the null value), the parameter value is declared to be “noninferior” to the null value (the same method can be used in the other direction, e.g., for “nonsuperior” inference).

Here we combine these methods in the following manner. If the CI does not overlap with ROPE, we deem the strength of evidence to be robust. If the CI does not overlap with the null value (usually zero), but partly overlaps with ROPE, we deem the strength of evidence to be moderate. If the CI overlaps with ROPE and the null value (in the manner usually described in noninferiority testing), we deem the strength of evidence to be modest. If the CI completely “spans” ROPE (i.e., overlaps and exceeds both ends of ROPE), we deem the evidence to be inconclusive. If the CI is contained within rope, we interpret the finding to support the null value. We deliberately follow these inferential criteria because it displaces the inferential focus from a binary “significant versus not-significant” label (Kruschke, 2018) to a more nuanced conclusion describing the “degree” of evidence in relation to meaningful benchmarks. We conducted a permutation analysis on several study variables to determine whether zero was an empirically reasonable null value to use. Analyses showed the expected value of a correlation coefficient under a permuted distribution was indeed zero. Posterior distributions were summarized with the mean of the relevant statistic (e.g., Cohen's d, regression coefficient) and an 89% highest density CI (HDI; all samples within the interval have a higher density than those outside it). We used 89% CIs upon the recommendation that wider intervals (e.g., 95%) are more sensitive to Monte Carlo sampling error (McElreath, 2015; Makowski et al., 2019).

Datasets

Study 2 was comprised of secondary analysis of three large existing datasets (N = 1,641 unique participants total) containing written text about parents and friends (nominated using the same procedure as described in Study 1) as well as social decisions involving these two others. We used one dataset (Dataset 1) to engineer a linguistic signature of social decision preferences. The written text data and social decision measures in this dataset were collected as part of a broader study aiming to answer an orthogonal research question (Guassi Moreira and Parkinson, 2024). The other two datasets (Dataset 2, Dataset 3) and additional social decision measures from Dataset 1 were used to validate the linguistic signature by ensuring that it predicted social decision preferences in additional contexts. These datasets come from a series of studies on social decision-making conducted by the research team (Guassi Moreira et al., 2018, 2020, 2021). These prior studies have yielded published results, but the written text data have not been analyzed before. A detailed description of each dataset follows.

Dataset 1 was comprised of five independently collected subsamples (S1.1–S1.5). Participants in each subsample were asked to nominate a parent and close friend of their choice and provide ∼10 adjectives or short phrases (e.g., “happy go lucky,” “pain in the butt”) describing each close other. Participants in S1.1 were also prompted to recall four types of memories involving each social partner—instances in which they felt supported by the close other, in which they supported their close other, in which their close other let them down, and their favorite memory with their close other. Social decision preferences were assessed with a one-shot dictator game question in which individuals allocated a hypothetical $100 pot between parents and friends in $5 increments and another in which they indicated who they would rather spend time with given a day with no other obligations or commitments. Subsamples S1.1–S1.5 totaled N = 870 participants (range of subsample size n = 45–454). Subsamples were recruited from the UCLA undergraduate psychology subject pool and online crowdsourcing platforms (Mturk, Prolific) in 2022. All prior analyses involving the social decision-making data from this dataset were recently published (Guassi Moreira and Parkinson, 2024).

Dataset 2 was comprised of five independently collected subsamples (S2.1–S2.5) as a part of a series of studies on social decision-making. Participants in each subsample were asked to nominate a parent and close friend of their choice, provide a memory they share with each target, and list several words and phrases describing the target. They were then administered a social decision-making task that forced them to make choices between the parent and friend. Social decision preferences were assessed with a version of the Columbia Card Task (Figner et al., 2009)—a multi-trial risk-taking paradigm—that was modified to capture social decision preferences when making choices with conflicting outcomes between a pair of others (Guassi Moreira et al., 2018). Subsamples S2.1–S2.5 totaled N = 574 participants (range of subsample size n = 46–225). Subsamples were recruited from the UCLA undergraduate psychology subject pool and the broader West Los Angeles community between 2016 and 2020.

Dataset 3 was comprised of four subsamples (S3.1–S3.4). Some of the participants in these subsamples also provided data for subsamples in Datasets 1 and 2. The distinctive contribution of this dataset relative to the others is that it uses a different social decision-making paradigm. As with the other datasets, participants were prompted to nominate a parent and close friend. Some subsamples followed the text generation procedure described for Dataset 1; others follow the procedure for Dataset 2, meaning that all subsamples provided adjectives and short phrases describing each other; and some also provided a memory of each other. Social decision preferences were assessed with a version of a multi-trial delay discounting task that was modified to capture social decision preferences when making choices with conflicting outcomes between a pair of others (Guassi Moreira et al., 2021; the same task as in Study 1). Subsamples S3.1–S3.4 totaled N = 920 participants (range of subsample size, n = 61–454). Subsamples were recruited from the UCLA undergraduate psychology subject pool (2018–2019) and from an online crowdsourcing platform (2022).

Further information about each dataset, such as participant demographics and methodological details, can be accessed in the original publications cited above. Self-reported relationship quality between participants and their nominated parents and friends was collected in all three datasets. All participants provided consent in accordance with the policies of the UCLA Institutional Review Board. Depending on their method of recruitment, participants were either compensated in course credit or monetary remuneration (USD). All study materials and code are publicly available on the OSF (osf.io/fzds6). For privacy reasons, we are unable to share the raw or preprocessed text data.

The following two sections detail (1) how we used Dataset 1 to engineer a linguistic signature of social decision preferences based on one social decision measure and then (2) how we used data from all three datasets to further validate the signature to verify that scores derived from the signature were related to social decision preferences in held-out data and held-out social decision measures.

Engineering a linguistic signature via machine learning

Since our goal was to examine whether linguistic representations of close others encode information about social decision preferences, we trained a model to predict social decisions from semantic features of written text data. We sought to produce a linguistic signature of social decision preferences that followed the logic of neural signatures and pattern expression analysis in the cognitive and affective neuroscience literature (Wager et al., 2013; Chang et al., 2015), both in the process used to engineer the signature and its application. The first step in this process was to test candidate models with cross-validation on the subsamples from Dataset 1 to settle on model specifications (type of model, hyperparameters). The pipeline for this task was comprised of the following steps: (1) social decision measure selection, (2) text preprocessing, (3) representation extraction, (4) leave-one-out cross–validation using each subsample, and (5) estimating model weights on the entire dataset using the model with the best cross-validation accuracy.

Selecting a social decision measure for signature construction

We used the forced-choice question about which close other the participant would rather spend a free day with (coded such that parent, 1; friend, 0) as our measure of social decision preference. We chose this metric for assessing social decision preferences for four reasons. First, the binary nature of the item makes for an intuitive classification metric with which to evaluate model performance (percentage accuracy). Second, using a binary forced-choice measure prevents participants from being impacted by a neutral scale response bias and requires them to disclose an unambiguous preference. Third, framing social decision preferences in terms of spending time (a finite resource subject to opportunity cost) with others simulates a real-world scenario that could generalize better to behaviors outside of the lab. Last, the simplicity of the item may help it better translate to other, more complex measures of social decision-making. Preferences for parents over friends for each sample are as follows: S1.1 = 51% chose parents over friends; S1.2 = 42%; S1.3 = 42%; S1.4 = 46%; S1.5 = 42%.

Text preprocessing

Written text data were preprocessed in a manner consistent with other similar studies (Fatima et al., 2021), implemented with the TextBlob python library. The following preprocessing pipeline was applied to each participant's data. We first removed stopwords, as defined by TextBlob, in addition to the words “favorite” and “memory” (as many participant responses to memory prompts began with this phrase). Afterward, each word was forced into its lowercase form and lemmatized. Words were then tagged as parts of speech and filtered. Only nouns, adjectives, verbs, and adverbs were in the preprocessed word list, corresponding to the following TextBlob tags: “JJ,” “JJR,” “JJS,” “NN,” “NNS,” “NNP,” “NNPS,” “RB,” “RBR,” “RBS,” “VB,” “VBD,” “VBG,” “VBP,” “VBZ.” Parent and friend memories were preprocessed separately, meaning that each participant had a set of a preprocessed words from written text data about their nominated parent and another preprocessed word set about their friend.

Representation extraction

We used word embeddings to extract linguistic representations of parents and friends from participants’ written text data. This approach quantified semantic meaning from the preprocessed word set in a bottom–up fashion using pretrained Word2Vec word embeddings. We extracted word embedding vectors for each word in the parent and friend word sets separately and then averaged the vectors of all set words together. We used the pretrained embeddings from Google's 300-dimension Word2Vec model (Mikolov et al., 2013), as implemented in the genism python package. Words in the word set that did not have a corresponding word embedding vector were excluded from the representation. In this approach, the semantic content of each participant's text data for a given close other (parent, friend) was summarized as a composite 300-dimensional vector. We used PCA to reduce dimensionality aid modeling. Both parent and friend composite vectors for each subject were reduced to 30 dimensions, consistent with similar work (Wang et al., 2023). On average, 30 dimensions accounted for ∼83% of the variance in written text data, collapsing across parent and friend text in all datasets.

Leave-one-subsample-out cross–validation

We used leave-one-subsample-out cross–validation to determine the best specifications for defining a linguistic signature of social decision preferences. We specifically sought to identify the best type of machine learning model and set of hyperparameters. We identified three possible types of machine learning models: a support vector classifier (SVC), a ridge regression classifier (RRC), and regularized logistic regression (RLR). Hyperparameters for each model were selected using a grid search within each iteration of cross-validation. The candidate values for all possible hyperparameters (C, γ, α) were {1, 5, 10, 50, 100}.

Our cross-validation approach took advantage of the independent subsamples in Dataset 1—we chose to leave one subsample out for each iteration of cross-validation. Since the goal of cross-validation is to estimate out-of-sample modeling performance, we reasoned the natural partitioning of each subsample as independent datasets best served this goal. This meant that each iteration of cross-validation proceeded such that one subsample was scaled and set aside, the other four were aggregated and scaled, a grid search was performed to find the best hyperparameter(s), the best hyperparameter was then used to estimate model weights on the four aggregated subsamples, and the model weights were finally applied to the left-out subsample to estimate the predictive accuracy of the model. This process was repeated five times, leaving one of the subsamples out each time, for all of the three models described above. This workflow was implemented using functions from the scikit-learn Python library.

Estimating model weights

As previously noted, the binary forced-choice question about spending time with either a parent (coded as 1) or friend (coded as 0) served as our criterion. The principal components derived from both the parent and friend text data (60 total) were entered into the winning model from the prior step as predictors. The ensuing 60 coefficients (weights) from the model indicate how a “preference” toward one close other is embedded in the linguistic representation and thus comprise the linguistic signature.

Validating the linguistic signature

Once we engineered in our linguistic signature in Dataset 1, we set out to further validate it by correlating linguistic preference scores derived from the signature with other measures of social decision-making across the three datasets. The logic for doing so is that if our linguistic signature is truly capturing social decision preferences between parents and friends, then the signature should generalize to other measures and predict preferences across different measures in the three datasets. To that end, we used the weights from our signature to compute preference scores and test whether those preference scores could predict social decision preferences on previously untested social decision tasks. Linguistic preference scores were calculated by taking the dot product between the model weights that comprised the linguistic signature and the feature-reduced text data from a given subject. The feature-reduced text data were the same as described above: PCA-reduced composite word2vec embeddings for parents and friends. When validating the linguistic signature, we used decisions from a one-shot dictator game where participants had to directly allocate money between parents and friend (Dataset 1), a multi-trial decision-making task that pitted gains and losses for parents against those for friends in a risky context (Dataset 2), and a multitrial decision-making task where participants had to choose to allocate money for, or time spent with, their parents and friends in a delay discounting context (Dataset 3; the same task used in Study 1). Finally, because social decision preferences are often predicted by attitudes toward social partners, we also tested whether the linguistic signature could predict relationship quality scores for parents and friends (available in all datasets). These validation tests for each outcome are described in greater detail below. Bayesian statistics were used for all analyses. The inferential criteria from Study 1 were applied to Study 2 analyses.

Social preferences in the one-shot dictator game

The one-shot dictator game, collected in Dataset 1, asked participants to allocate a hypothetical pot of $100 between their nominated parent and friend. Values varied by increments of $5, ranging from ($100 parent, $0 friend) to ($0 parent, $100 friend), resulting in a 21-point ordinal variable. Responses were recoded to be the percentage of money allocated toward the parent (0, 0.05, 0.10, …, 0.90, 0.95, 100). A robust Pearson's correlation between preference scores and dictator game allocations was computed using the Bayesian software package stan. We note that while the text data for this validation step are not novel (i.e., were used in model engineering), the outcome variable (dictator game allocations) is, and it is valuable to know whether preference scores can predict decision preferences in different contexts within-participants.

Social preferences in a risky context

Linguistic preference scores were next validated by determining whether they could predict parent versus friend preferences in a risky decision context. Participants in this dataset (Dataset 2) completed a multi-trial decision task that required them to make choices between a risky and a safe outcome. The risky outcome carried the potential for either a hypothetical monetary gain or a hypothetical monetary loss according to a known probability; the safe outcome was neutral (no gain or loss). The interests of parents and friends were pitted against each other such that any potential gain from a risky choice on a given trial would be credited to one partner, whereas any potential loss would be credited to the other partner. Participants completed multiple trials across two runs of the task, one in which risky gains benefitted one partner (e.g., parent), while the losses were incurred by the other partner (e.g., friend), and another run in which the opposite was true. Hierarchical Bayesian logistic regression was used to model the trial-level likelihood of making a risky (vs safe) decision as a function of a parameter that encoded choice preferences on the task and other trial-level covariates. We refer readers to Guassi Moreira et al. (2018) for details about the form of the within-person component of the model. We entered preference scores derived from the linguistic signature as a between-person predictor to the model, allowing it to interact with the choice preference parameter from the model. Because the choice preference parameter was allowed to vary randomly across subjects, the ensuing interaction term codes for how choice preferences on the task vary as a function of preference scores derived from the linguistic signature. The brms package was used to fit the models (four chains, 1,000 discarded burn-in samples, 2,000 samples/chain, no thinning).

Social preferences in a delay discounting context

Linguistic preference scores were further validated by testing whether they could predict parent versus friend preferences in a delay discounting context. This is the same task used in Study 1, with the only difference being that more trials were administered to participants. Hierarchical Bayesian logistic regression was used to model the trial-level likelihood of making a discounting decision (i.e., choosing the immediate over delayed reward) as a function a parameter that encoded choice preferences on the task and another trial-level covariate (same within-person model as Study 1). We ran separate models for choices with monetary and social outcomes. Preference scores derived from the linguistic signature were entered into this model in the same manner described above. The brms package was also used to fit the models (using the same specifications that are listed above).

Predicting relationship quality using the linguistic signature

Last, we opted to relate preference scores derived from the linguistic signature to self-reported relationship quality with parents and friends. In doing so, our goal was to subject the linguistic signature to another validatory step that involved predicting “attitudes” that may also influence social decision behavior. A meaningful result for this test would indicate that motivationally relevant information about one's attitude toward a close other is encoded in their linguistic representation. Information about the self-report measure used here (Inventory of Parent and Peer Attachment) and its reliability is included in the Extended Data.