Dynamic time warping of interpersonal physiologic synchrony predicts high stakes trust

We trust strangers every day. Deciding to trust another depends on the social and legal environments that influence calculative trust as well as unconsciously processed neural signals. Unenforceable communication (“cheap talk”) has been shown to increase cooperative behaviors, but why it does this is not well-understood. Herein we test whether physiologic synchrony during pre-decision communication explains why one would trust a stranger when the money at risk is meaningful. Working-age adults discussed how to share $480 before making sequential decisions in private. The social interaction was structured so that the first decision-maker had to sacrifice money s/he controlled by sending it to the second decision-maker in order to grow the money at stake. Electrodermal activity was captured during communication and two measures of synchrony, intersubject correlation and dynamic time warping, were evaluated for their predictive accuracy. We found that the dynamic time warping measure of synchrony predicted trust with a stranger with 60% accuracy, a 15% improvement over the accuracy using intersubject correlation. Our findings demonstrate that dynamic time warping is a more effective measure of physiologic synchrony, leading to new insights into the mechanisms through which communication influences the decision to trust a stranger.

Trust is extraordinarily important in building personal relationships and reduces frictions in economic exchange, being a potent engine for economic growth (Algan and Cahuc, 2010; Zak, 2012; Algan and Cahuc, 2014; Zak and Knack, 2001). Neuroscience researchers have made significant progress understanding why people are trustworthy after they have been trusted, identifying a key role for the neurochemical oxytocin (Kosfeld et al., 2005; Zak et al., 2004; 2005). Being trusted by another also reduces stress hormones, increasing the likelihood of reciprocation (Curry et al., 2015; Morhenn et al., 2008; Morhenn et al., 2012). Yet, why people initiate trust, especially when the stakes are high, is poorly understood (Hancock et al., 2023; van der Werff et al., 2019). Personality traits such as agreeableness, conscientiousness, and openness to experiences play a role, but do not explain a substantial portion of the decision to trust (Freitag and Bauer, 2016; Mooradian et al., 2006). The brain hides neurochemical signals and associated neuroelectrical activity from conscious awareness, so asking participants why they trusted a stranger yields little insight (Cheshire et al., 2010; Ostrom and Walker, 2003).

The lack of understanding of why people initiate trust in laboratory experiments, even while trust is a common occurrence, may be an artifact of experimental design. The first trust experiments used private, sequential monetary transfers in order to reduce intersubject variation that could influence decisions (Berg et al., 1995). These studies disallowed pre-decision communication and most subsequent trust experiments have followed the same approach. Communication that is not binding, known as “cheap talk” in game theory (Farrell and Rabin, 1996), was thought to add external factors such as attractiveness, inferred income, smiles, touch, or ethnicity that could affect decisions (Andreoni and Petrie, 2008; Gereke et al., 2018; Morhenn et al., 2008; Wilson and Eckel, 2006). Yet, the behavioral data show that cheap talk substantially increases trust (Ben-Ner et al., 2011; Johnson and Mislin, 2011; Lev-On et al., 2010). Indeed, outside the lab, trust is nearly always preceded by communication during which the terms of exchange are discussed. By removing communication, these studies have ignored the most interesting aspect of the decision to trust a stranger.

Another artifact that inhibits the understanding of trust is the small stakes common in most trust experiments (Johnson and Mislin, 2011; Smith, 1998). Ernest Hemingway wrote that, “The best way to find out if you can trust somebody is to trust them.” Yet, this carries substantial risk when the stakes are high. Low stakes monetary trust experiments typically yield 90% of individuals trusting strangers (Camerer, 2011; Zak, 2012). The use of high-stakes, as is done here, ads weight to the trust decision by providing an incentive not to trust and instead keep an initial endowment. In addition, high stakes trust experiments provide evidence that is more likely to generalize to behaviors outside the lab where decisions that depend on trust often have consequential outcomes (Kunreuther et al, 2002; Zak et al., 2022).

In addition to small stakes, the majority of trust experiments use a convenience sample of college students. This also inhibits generalization since most college students are WEIRDs (Western, Educated, Industrialized, Rich, and Democratic; Henrich, 2020). Students behave atypically in many experiments compared to the general population (Henrich et al., 2010). This is compounded by observing other college students entering an experiment which leads to an assumption that they have similar backgrounds and will behave in ways similar to oneself, increasing the likelihood of trust (Chuah et al., 2023; Finseraas et al., 2019).

The present experiment added four aspects that are rare in most trust studies in order to increase ecological validity and identify why people initiate trust. First, we only recruited demographically diverse adults aged 25–50 to participate. Second, participants could earn up to $530 based on one's decision and that of the participant to which s/he was matched. Third, after all participants were instructed in the monetary trust task, they had two minutes to meet their counterparts and discuss what they intended to do prior to making private decisions. The fourth innovation was to capture physiologic activity during the communication period to seek to identify a mechanism that influenced the decision to trust a stranger.

We hypothesized that physiologic synchrony during the communication period would increase the likelihood of trusting a stranger in the monetary decision task. Our analysis used two measures of physiologic synchrony, one new to the behavioral neuroscience literature, dynamic time warping discussed in detail below, to see which one produced the most accurate prediction of the trust decision. The choice of which neurophysiologic measure to use, where our use of “neurophysiologic” includes both central and peripheral neurologic responses, was driven by both its likely predictive ability and a desire to minimize participant interference in order to maintain ecological validity. The present study measured electrodermal activity (EDA) during the communication period as this measure has been shown to predict why people renege on promises to reciprocate trust and is straightforward to obtain (Zak, et al., 2022).

Physiologic synchrony measures

Early studies of physiologic synchrony found positive correlations for EDA and heart rate while patients spoke to therapists (Balconi and Fronda, 2022; Coleman et al., 1956; Di Mascio et al., 1955; Stuldreher et al., 2020) and the findings appear to be robust (Golland et al., 2015). EDA captures the peripheral physiologic response of centrally-mediated arousal through increased conductivity from sweat in the eccrine glands on the hands (Critchley, 2002; Figner and Murphy, 2011).

Later research found that neurophysiologic synchrony occurs during cooperative and ritualized behaviors (Jackson et al., 2018; Xygalatas, et al., 2013). Synchrony correlates with a number of psychosocial constructs, including shared attention, empathy, attachment, conflict, facial mimicry, vocal prosody matching, and emotional co-regulation (Jospe et al., 2020; Mayo and Gordon, 2020; Palumbo et al., 2017). While the mechanisms producing physiologic synchrony have not been fully characterized, this effect appears to be facilitated by both conscious and unconscious communication of information about another's intentions and/or emotional states (Wood et al., 2021).

Studies of strategic decision-making have begun to examine the effects of physiologic synchrony on cooperative behaviors. One study directed participants to build model cars with LEGOs after which they could contribute money to a common investment pool that would be split equally between participants. Cardiac synchrony was correlated with the expectations of others' contributions to the common pool although not contributions per se (Mitkidis et al., 2015). Synchrony was similarly measured in iterated prisoner's dilemma games in which treatment participants could see their dyadic partners while controls could not (Behrens, et al., 2020). A time-lagged cross correlation measure of the synchrony of participants' EDA was associated with cooperation, but only when participants could see each other.

There are a number of measures of physiologic synchrony (Palumbo et al., 2017). One used in neuroimaging and electroencephalogram (EEG) studies is intersubject correlation (Golland et al., 2014; Nastase et al., 2019). Intersubject correlation (ISC) measures synchrony by calculating Pearson correlations for a physiologic measure between participant pairs (Boker et al., 2002). ISC, as well as other measures of physiologic synchrony, often use a time lag to account for the dynamics of interpersonal interactions (Boker et al., 2002; Schoenherr et al., 2019). This approach is known as windowed cross-lagged correlations (WCLC) and uses four parameters: window size, window increment, maximum lag, and lag increment. Time series are broken into short, overlapping windows that are used to find moving estimates of association (Behrens et al., 2020). Overlapping windows also prevent missing points of synchronization that may occur at the edges of adjacent segments. For each window, segments are lagged up to a given maximum that accounts for the potential delay between physiologic responses (Behrens et al., 2020). ISC is a traditional measure of synchrony so we use it in our analysis as a benchmark.

We have also included a second measure of physiologic synchrony, dynamic time warping (DTW), that was developed to align time series that vary in speed and is thus well-suited to responses during dyadic human communication (Kruskal, 1983). Indeed, one of the first and well-known applications of DTW was speech recognition, suggesting it would be an appropriate measure of synchrony for EDA signals during communication (Juang, 1984; Permanasari et al., 2019). Other applications include alignment of video components, computer vision and animation, surveillance, protein sequence alignment, music and related signal processing, robotics, and manufacturing (Senin, 2008). DTW has been applied to physiology to assess abnormal and diseased cardiac responses (Shorten and Burke, 2014) and, as we do here, and to EDA while people watch films (Kostoulas et al., 2017). As with ISC, DTW uses windows that are based on an integer absolute difference between two time series, known as the Sakoe–Chiba window (Sakoe and Chiba, 1978). Various window lengths are used to align data, including physiologic data, that occurs between participants. A zero window aligns each time series temporally, while larger windows offset time series by an integer amount applied to the frequency of data collection.

DTW is a more comprehensive measure of synchrony than ISC and thus a goal of the present research was to test whether DTW will more accurately predict interpersonal trust when compared to ISC. For example, most physiologic time series are nonstationary and DTW analyses do not require stationary time-series data (Dilmi et al., 2020). While ISC does not require stationarity either, statistical tests using nonstationary ISC are misspecified (Nastase et al., 2019). Since the DTW technique may be new to some, a primer on DTW is included in Appendix A.

Procedure

This experiment was conducted at the Center for Neuroeconomics Studies and was approved by the Institutional Review Boards of Claremont Graduate University (#1006) and the United States Department of the Air Force (#FWR20110168X) and followed the guidelines of the Declaration of Helsinki. Participants gave written informed consent prior to inclusion and were assigned an identity-making code. There was no deception in any part of this study and participants earned $50 as an incentive to show up. After providing written informed consent, participants were led to partitioned computer stations where they were fitted with disposable Ag-AgCl electrodes placed on the palmar surface of the middle phalanges of the index and middle fingers of their non-dominant hands for EDA data collection. While waiting for others to be fitted with electrodes, demographic data were collected from participants.

Trust task

Participants were fully and identically instructed in a sequential decision task known as the trust game (TG; Berg et al., 1995). Participants were randomly assigned as the first decision-maker (DM1) or second decision-maker (DM2). All decisions were made in private at computer terminals but were preceded by a communication period (see below). Our version of the TG endowed both dyadic participants with $120. DM1 then was prompted by computer software to send either $0, $40, $80, or $120 of his or her endowment to DM2 electronically. Both DMs were informed that transfers sent by DM1 would be removed from DM1's account and would be multiplied in DM2's account as follows: $40 turns into $240, $80 turns into $360, and $120 turns into $480. The typical TG uses the same multiplier applied to DM1 to DM2 transfers; here we sought to make the decision process more complicated to induce more consideration of the outcomes by varying the multiplier. This approach was also used to ensure participants understood that the amounts involved were meaningful (Zak et al., 2022).

DM2s were next informed via software of the multiplied amount received from DM1s and the total in their accounts (the multiplied transfer plus the $120 endowment). Then, each DM2 was prompted by software to choose a return transfer to the DM1 in his/her dyad. Both DMs were informed that return transfers would be removed from DM2s' accounts and would be transferred without multiplication into DM1s' accounts. Participants received printed tables showing all possible decisions and how these affected their own, and their dyadic partner's, earnings (Figure 1). As shown in the figure, when DM2s received nonzero transfers from DM1s, they could return nothing ($0) maximally benefiting themselves or could split the multiplied stake with DM1s. The consensus view in experimental economics is that the DM1 to DM2 transfer measures trust, while the DM2 to DM1 return transfer measures reciprocity or trustworthiness (Smith, 1998). As a control for altruism (Cox 2004), we also asked participants to make a unitary monetary transfer decision from an endowment of $120 to their dyadic partner prior to the TG decision. This decision task, known as the dictator game (DG), is typically used to measure altruism (Camerer, 2011). To reduce the effect of baseline altruism, the following analyses removed the DG transfer from the DM1 transfer as in previous research (Cox, 2004).

Communication

After instruction in the DG and TG, but prior to decisions, dyads were seated facing each other in a private room for two minutes and were advised to discuss their strategies and make a commitment to a particular choice. Participants were informed that these commitments were not enforceable and decisions would be made in private in partitioned computer stations. EDA was measured throughout the communication period.

Physiology

EDA data were collected with a Biopac MP150 (Biopac Systems, Inc., Goleta, CA). After fitting, research assistants performed an impedance check to ensure signal quality, with impedance continuously monitored during the experiment. Baseline EDA was obtained by having participants sit at partitioned stations with headphones on for two minutes. Each participant's baseline EDA response was removed from that occurring during communication to control for individual differences as is standard (Boucsein, 2012).

After collection, EDA data were cleaned using the AcqKnowledge software package (Biopac Systems Inc., Goleta, California). Specifically, the data were converted to microSiemens (μ S) and, to correct for positive skew, were square root transformed (Braithwaite et al., 2013). Periods of excessive noise or signal drop were linearly interpolated as in previous analyses using an automated algorithm (Johannsen and Zak, 2020). Then, a 10-Hz low-pass filter was applied to remove high-frequency noise (Norris et al., 2007). Finally, all data were visually inspected to ensure that the automatic data cleaning identified and corrected artifacts. After ensuring data quality, the change in skin conductance levels (SCL) was calculated during the 2-min communication period by subtracting baseline SCL. SCL is a tonic measure of electrical conductivity and is among the most common EDA measures (Figner and Murphy, 2011).

Participants

The data were collected as part of a broader neurophysiological study (Willoughby et al., 2012; Zak et al., 2022) that was designed to be ecologically valid by studying an ethnically diverse sample of working-age adults making decisions using meaningful amounts of money. The participant pool consisted of 76 employed adults between the ages of 25–50 (32 females, M Age = 30, S D Age = 5.73) from diverse cultural backgrounds. Participants self-identified as White (47%), Asian (20%), Latino/Hispanic (16%), and African American/Black (17%). Participants were a mix of native-English speakers born and raised in the United States (46%) and non-native English speakers born and raised outside the United States. Participants were recruited via the Center for Neuroeconomics Studies website and by direct outreach to targeted community groups. Participants were pseudo-randomly paired, resulting in 47% mixed-gender pairings. All participants made four decisions in the TG preceded by four communication periods with dyadic partners. Data loss occurred due to sensor slippage, signal interference and participant fatigue resulting in 222 interactions.

Statistical analysis

The EDA data were down sampled from 250 Hz to 1 Hz to reduce noise. Then, ISCs were calculated to measure physiologic synchrony during the communication period using window sizes 2, 3, 5, and 8, each with a lag of 1. There is no established theory on window size, so several were investigated to assess best fit (Gates and Liu, 2016; Schoenherr et al., 2019). DTW was also calculated for windows of 0, 5, 10 and 15. ISCs and DTWs for each window size were tested for their ability to rank the four trust groups (none, low, medium, and high trust) for physiologic synchrony. Shapiro–Wilk tests (see below) showed the ISC and DTW data were not normally distributed, so the nonparametric Kruskal–Wallis test was used to compare differences. Post-hoc non-parametric pairwise comparisons were done using Dunn's test which is based on rank sums and Spearman's rank correlation coefficient (denoted ρ) and used Bonferroni corrected p -values (Dinno, 2015). In order to compare the accuracy of ISC and DTW, binary logistic regressions were estimated using all windows for both measures of synchrony. The high correlation between measures (.49–.99) resulted in multicollinearity. To resolve this, principal components (PCs) were separately generated from both types of synchrony measures using cross-validation to choose the optimal number of PCs (Bro and Smilde, 2014). The use of a logistic regression was included to demonstrate the practical use of physiologic synchrony in predicting behavior while ensuring that regression requirements were met (Shrestha, 2020). Logistic regression accuracy will be reported using precision (true positive rate), recall (true negative rate), balanced accuracy (average of precision and recall), area under the curve (AUC) for the receiver-operator curve, and F1 (harmonic mean of precision and recall).

Data availability

The data are posted at Open ICPSR openicpsr-238643.

Behavior

Across all sessions, maximal trust in which DM1s sent their entire $120 endowment to their dyadic partners occurred in 62 interactions (28%). Conversely, in 89 interactions (40%), DM1s sent $0, revealing complete distrust. The decision to trust another person did not vary by race (p =.154), sex (p =.103), age (p =.201), education (p =.069), or income (p =.075), nor did it depend on whether dyadic partners were the same or different sexes (p =.644). The amount of the DM1 transfer was uncorrelated with age (r = −.005, p =.938), education (r =.0583, p =.401), or income (r =.055, p =.429). Figure 2 shows the distributions of amounts DM1s sent to their partners.

Skin conductance levels

The change in SCL (ΔSCL) from baseline to the communication period did not vary between those who did not trust others and those who did (M trust =.10, SD = 0.183; M notrust =.106, SD =.20; t = −.241, p =.810). There was also no difference in ΔSCL during communication for low trust (sent $40 or $80) and maximum trust (sent $120) (M low trust =.116, SD =.176; M max trust =.081, SD =.191; t = 1.071, p =.286).

Intersubject correlations

The ISCs averaged across the four window sizes while dyadic pairs were communicating did not vary between trust vs. no trust (M trust =.671; M notrust =.680; t = −.667, p =.505) or by the degree of trust (M low trust =.665; M max trust =.677; t = −.639, p =.524).

Next, Shapiro–Wilk tests were performed on each window size which showed that the data were not normally distributed (window size 2 (W2), lag 1, W =.98, p =.02; window size 3 (W3), lag 1, W =.93, p <.001; window size 5 (W5), lag 1, W =.96, p <.001; window size 8 (W8), lag 1, W =.95, p <.001). As a result, the nonparametric Kruskal–Wallis test was used to compare ranked ISCs. None of the ISC values were significantly different from the others (W2: H (3) = 6.66, p =.08; W3: H (3) = 1.00, p =.80; window size 5: H (3) = 3.41, p =.33; window size 8: H (3) = 3.00, p =.39; Figure 3).

Dynamic time warping

Similar to the ISC results, DTW values averaged across the four windows were identical for all levels of trust compared to no trust (M trust = 88.077; M notrust = 85.236; t =.642, p =.522) but were statistically different by the degree of trust (M low trust = 95.346; M max trust = 79.752; t =: 2.887, p = 0.005).

Shapiro–Wilk tests were performed on each window width and showed that two of the four windows were not normally distributed (W0: W =.96, p <.001; W5: W =.99, p =.44; W10: W =.99, p =.12; W15: W =.98, p =.005). As above, nonparametric Kruskal–Wallis tests were used to compare the four groups. Three of the four windows were significantly different from each other, windows 5, 10, and 15, but not window 0 (W0: H (3) = 6.61, p =.09; W5: H (3) = 9.14, p =.03 W10: H (3) = 9.39, p =.02; W15: H (3) = 8.52, p =.04). The relationship of each DTW window to trust was examined using Dunn's test with Bonferroni corrected p-values. Consistent with our hypothesis, DTW synchrony during communication was associated with increased DM1 trust for all four windows tested, i.e., 0, 5, 10, and 15 (ps <.01; Figure 4).

Predicting trust

We estimated binary logistic regressions (trust or no trust) to compare the predictive accuracy of the ISC measure of synchrony with the DTW measure to further assess which approach is superior. As discussed in Methods, all four windows could not be included in regressions due to their high correlations. Instead, we generated principal components separately for the ISC and DTW measures from the four windows in order to quantify predictive accuracy.

The ISC analysis generated two PCs which explained 83% of the variation in synchrony. Neither variable was statistically associated with the decision to trust a stranger in a logistic regression (ps >.08, Appendix Table A1). The regression had a balanced accuracy of 52% (precision =.63, recall =.38, AUC =.54, F1 =.47, VIFs ∼1). Re-estimating the logistic regression including age and sex as controls produced an accuracy of 50% due to missing values (precision =.59, recall =.55, AUC =.49, F1 =.57, VIFs ∼1).

We next generated PCs for the DTW measures and used them to estimate a logistic regression to establish DTW's relationship to trust. The principal components analysis produced three PCs which explained 99% of the variation. None of these variables alone was statistically associated with the decision to trust a stranger (ps >.14, Appendix Table A2). The regression estimation using the three PCs produced a balanced accuracy of 60% (precision =.68, recall =.68, AUC =.58, F1 =.68; VIFs ∼1). Including demographic covariates in the logistic regression (ps >.23) dropped the accuracy of 56% due to missing observations (precision =.70, recall =.35, AUC =.55, F1 =.47, VIFs ∼1).

Why trust a stranger? Nobel laureate Vernon Smith has said that, “the trust games …were a puzzle: why were people so cooperative?” (Schrepel, 2023). Why people reciprocate in trust games is now fairly well-understood due to the release of the neurochemical oxytocin (Zak, 2012; Zak et al., 2004; 2005). But, why one would initiate trust with a stranger is poorly understood, though personality traits and attitudes play a role (Grünhage and Reuter, 2022; Hancock et al., 2023; Stavrova et al., 2023; Thoelen and Zak, 2024; Zak et al., 2005). Such decisions are especially important to understand when significant resources are involved (Ariely et al., 2009; Hersch and McDougall, 1997; Zak et al., 2022).

The present study modified the traditional trust game in four ways in order to generate findings that could generalize outside of the lab. First, we used monetarily meaningful stakes wherein participants could earn up to $530. Second, decisions were made by working-age adults rather than a convenience sample of students. Third, decision-makers met each other in person to discuss their choices and could make verbal commitments prior to executing private decisions. Fourth, we measured EDA during communication to capture a neurophysiologic synchrony.

We hypothesized that physiologic synchrony during the communication period would demonstrate why DM1s sent money to DM2s and that DTW would be a better measure of synchrony than ISC. Our results support both hypotheses and provide evidence that physiologic synchrony during communication influences consequential decisions. These results provide new insights into choices that people regularly face: should I trust a stranger or not?

Even though pre-decision “cheap talk” is not enforceable, it does reveal information that influences people's choices. This information includes the statements people make (Farrell and Rabin, 1996), facial expressions such as smiles (Reed et al., 2018), word choice (Lee et al., 2021), gestures and other types of bodily movements (Wang et al., 2017), signs of physiologic stress which indicate that people are lying (Ten Brinke et al., 2019; Zak et al., 2022), and, as we have shown here, physiologic synchrony. Much of the information we obtain from other people is processed unconsciously, necessitating the use of technology to measure it. Even though people are generally not consciously aware of neurophysiologic responses, they still influence behavior (Bargh, 2014).

The unconscious signals we identified that the influence the choice to trust a stranger with a meaningful amount money were part of a larger study in which multiple moderately-invasive neurologic measures were captured in order to develop a comprehensive understanding of the initiation of trust (Zak et al., 2022). The instructions to participants made clear that they would earn money to compensate them for their time and discomfort and that all decisions would be paid in cash without any deception. That participants in trust games earn the money from their decisions is important in motivating attention to the choice set (Alos-Ferrer and Farolfi, 2019). The use of a meaningful amount of money may explain why nearly half of the participants in the present study showed complete distrust and one-third displayed maximum trust, a distribution not typically seen in other trust games (Hardin, 2001; Zak et al., 2005).

Part of this research program involves searching for observational “tells” that could help negotiators and others who are bargaining, to form expectations of another's intentions without the use of physiologic measurements. After identifying the neurophysiologic signals that predict trust, these can be mapped into observable movements, statements, and language. We introduced DTW to test whether it was a more effective measure of physiologic synchrony than the more commonly-used ISC. The analysis showed that applying DTW to EDA data increased predictive accuracy of the initiation of trust by eight percentage points when compared to ISC. While a modest improvement, the ISC measures were largely unrelated to the trust decision while the DTW windows were statistically associated with trust, demonstrating the overall value of using DTW to measure synchrony. Nevertheless, there are likely other variables that should be investigated that, along with EDA synchrony, contribute to the decision to trust a stranger. These may include touch, smiles, gender, personality traits, as well as participant income, education, and ethnicity (Cetre et al., 2024; Zak et al., 2005; Zak, 2012; Zak, 2017). Relating these variables to physiologic synchrony is reserved for future research.

We note that the DTW technique is not new, but it is only recently that it has been applied to neurophysiologic data (Krotov et al., 2025; Philips et al., 2022; Wiafe et al., 2024). Herein, we applied DTW to EDA data in order to predict subsequent interpersonal behaviors, which is a new application of this technique. We also recognize that ISC is a reasonable measure of synchrony and may be appropriate to explain some behaviors and for some types of neurophysiologic data, for example, with EEG or fMRI (Dinov et al., 2016; Jing et al., 2016; Meszlényi et al, 2017). Yet, DTW is particularly suited to speech data which generated the EDA signals we analyzed here (Permanasari et al., 2019; Silbert et al., 2014). DTW techniques have also informed behavioral studies of physical rehabilitation (Su, 2013), declining handwriting abilities (Di Brina et al., 2008), speech styles (Shahin and Botros, 1998), and the dynamics of depressive symptoms (Hebbrecht et al., 2020). Yet, we were unable to find any studies in the published literature that applied DTW to physiologic data in order to predict behavior as we do here. Thus, our contribution is two-fold: to introduce DTW as a methodology to explain interpersonal interactions, and to shed light on why people decide to trust, or not to trust, others, including strangers.

Yet, we are cautious not to generalize our findings too far. While we sought to create an ecologically-valid study, the sample size was moderate and participants were educated, Western adults. At the same time, EDA signals emanate from the brainstem and are conserved across individuals, suggesting our findings may hold on average for different populations and in different settings even while individual variations in EDA responses are large (Tronstad et al., 2022). Our findings should therefore be replicated and extended to ensure they do indeed generalize and should also assess the effect of phasic skin conductance responses. We also caution that the TG is a scripted social interaction that only partially reflects how individuals interact outside the lab. Future research should examine different populations in field studies using non-scripted tasks that quantify trust in order to provide further insights into why people trust strangers (Gao et al., 2025; Mayer et al., 2011). We would advocate, though, that to be of most use to scholars and society in general, future studies should include meaningful stakes, pre-decision communication, and should capture neurophysiologic signatures that reveal the drivers of decisions.