Some scholars view trust as an innate psychological disposition that is rigid (Nannestad, 2008; Van Lange et al., 2014; Dinesen and Bekkers, 2017) or fixed in childhood and not affected by day-to-day experiences later in life (Bowlby, 1969; Couch and Jones, 1997; Uslaner, 1999, 2002). Others believe that trust shifts daily or is impacted by life events and experiences (Rotter, 1971; Hardin, 2002; Glanville and Paxton, 2007; Paxton and Glanville, 2015). Supporting the latter perspective, trust has been found to increase with a breadth of foreign travel experience (Cao et al., 2014) and decrease after job loss (Laurence, 2015) or when frequently moving between neighborhoods within a 5-year period (Helliwell and Wang, 2011). First-generation immigrants moving from a less trusting to a more trusting country display higher levels of trust than do individuals in their countries of origin (Dinesen and Bekkers, 2017). As a whole, it seems likely that generalized trust has both malleable and stable components (Van Lange, 2015; Dinesen and Bekkers, 2017).

Studies across a range of fields indicate numerous individual- and community-level benefits associated with higher levels of trust. At the individual level, trust is positively associated with traits such as social and emotional intelligence, self-esteem, sense of control, attachment security, optimism, tolerance, and acceptance of dissimilar others, as well as lower suspiciousness, jealousy, and shyness (Yamagishi, 2001; Hooghe et al., 2012; Oskarsson et al., 2012; Carl and Billari, 2014; Dinesen and Bekkers, 2017). Trust also predicts a variety of prosocial behaviors, such as civic engagement, volunteering, and integration into one's neighborhood (Marschall and Stolle, 2004; Helliwell and Wang, 2011; Dinesen and Bekkers, 2017). Personal relationships are positively impacted by higher levels of trust (e.g., Larzelere and Huston, 1980; Simpson, 2007; Murray et al., 2012). In contrast, distrust is associated with poorly differentiated self-concepts, loneliness, social phobia, and Machiavellianism (Couch and Jones, 1997; Simpson, 2007; Rotenberg et al., 2010). Low levels of trust also predict depression, anxiety, and poor physical health (Bienvenu et al., 2004; Helliwell and Wang, 2011; Qualter et al., 2013; Widiger, 2015).

Several studies have considered trust differences across demographic groups. Croson and Buchan (1999) showed that women tend to be less trusting than men, despite women being more trusted (Cappelen et al., 2020). The positive relationship between age and trust has been shown to be robust across several measures of trust and holds across countries (Li and Fung, 2013). It has been suggested that this could be associated with prioritizing social and emotional connectedness later in life or even physical and cognitive decline, as older people tend to rely more on others (Li and Fung, 2013). A computational linguistics framework could reveal these signals through mentions of family, support, events, and even physical/cognitive decline.

At the community level, trust features prominently in several definitions of social capital. Putnam defined social capital as “social networks and the associated norms of reciprocity and trustworthiness” (Putnam, 2007, p. 137; see also Stolle et al., 2008). Trust has been measured as a facet of social capital by the social trust question (Claibourn and Martin, 2000; Grootaert et al., 2004), and there are large differences in social trust across different countries (Halpern, 2005) and states within the United States (Fairbrother and Martin, 2013).

Other work has considered differences in trust across cultural and ethnic groups, comparing individualistic and collectivistic cultures (e.g., Bond and Cheung, 1983; Yuki et al., 2005). At these larger aggregated levels, trust has been named a crucial social attribute in an increasingly globalized world economy, where interactions with unfamiliar others are inevitable (Dinesen and Bekkers, 2017). It is highly predictive of economic growth and more effective governments (Knack and Zak, 2003; Delhey and Newton, 2005), civic engagement (Uslaner and Brown, 2005), and cooperation with strangers (Dinesen and Bekkers, 2017). Trust further correlates with wellbeing and lower rates of suicide (Helliwell and Wang, 2011).

Social, political, ethnic, and economic diversity have been found to decrease generalized trust in some studies—although this effect is moderated by the quantity and quality of social interactions across these intra-community differences (Stolle et al., 2008). One might expect increased immigration into European countries to predict decreases in trust, but a study aiming to measure this relationship found no evidence (Hooghe et al., 2009). Other studies have shown that while immigration can reduce trust, the effect is substantially smaller than other factors, such as economic difficulties and levels of social connectedness (Sturgis et al., 2011). The dramatic shifts in trust in the United States and abroad noted at the beginning of this article may be occurring in response to an increasingly diverse population, or other factors could be at work. Researchers investigating these domestic and international shifts in generalized trust would be greatly aided by the ability to measure trust at larger scales and in shorter time intervals.

The relationship between trust and politics has been shown to be complex and contested. Van Ingen and Bekkers (2015) showed very small or non-significant causal effects between generalized trust and civic engagement despite the two being highly correlated. Although the correlations are robust, work has suggested that trust is not the direct mechanism behind increased engagement, but the increase of both trust and civic engagement seems due to an increase in underlying prosociality (Van Ingen and Bekkers, 2015). Others have found that trust (political and social), at both the individual and community levels, is associated with an increased likelihood of voting and voter turnout (Grönlund and Setälä, 2007; Rahn et al., 2009). It has also been shown that increased welfare programs and social spending are linked to higher levels of trust (Brewer et al., 2014). Other aspects of politics, such as partisanship, also have multifaceted relationships with trust. Hooghe and Oser (2017) showed that partisanship (strength of party affiliation) is positively related to political trust (i.e., trust in political institutions) but negatively linked to generalized trust. This relationship between political trust and partisanship is independent of party affiliation (e.g., Democrats and Republicans) and has remained stable over time. The negative relationship between generalized trust and partisanship shows that partisanship decreases social cohesion, which is in line with literature showing that increased party divisions contribute to polarization. Several of these studies examined trust across larger spatial units, such as countries, and thus, a finer grained county-level analysis may add additional context to this line of research.

Characterizing traits at the regional level is an emerging trend made possible by access to larger data sets and computational techniques. Regional personality traits, measured using a large-scale online survey, were found to have strong associations with social involvement, demographics, crime, religiosity, occupation, and health-promoting behaviors (Rentfrow et al., 2008; Ebert et al., 2022; Giorgi et al., 2022). In the subfield sometimes referred to as “geographical psychology,” regions are thought to differ, in aggregate, across personality and other characteristics based on selective migration patterns. That is, individuals often move to areas in which individuals better match one's own traits. In some cases, individuals can collectively develop characteristics that are advantageous in a given physical environment (e.g., climate, population density, and the availability of resources; Rentfrow et al., 2008). Social influence—that is, assimilating to the traits that are valued in one's region—and environmental influence may also lead to “regional traits.”

Trust has been primarily assessed through experimental observations, self-report scales, and single items in large national surveys (Nannestad, 2008). Lab-based experiments typically involve a task in which individuals can decide whether or not to cooperate (Berg et al., 1995). An illustrative example of experimental research in the field involves the dropped-wallet paradigm, in which a wallet containing personally identifying information and money is dropped in a public place and the rates of wallet return serve as a behavioral measure of trust in communities (Nannestad, 2008; Helliwell and Wang, 2011).

A variety of generalized trust scales exist, such as the Faith in People scale (Rosenberg, 1957) or Yamagishi and Yamagishi's (1994) Trust Scale. These scales, however, have been criticized for including factors of other constructs (e.g., optimism), combining assessments of one's own perceived trustworthiness with the trustworthiness of others, and including related but distinct aspects, such as trust in another's competence (Nannestad, 2008; Dinesen and Bekkers, 2017). Several studies have successfully used the trust items from the Agreeableness factor of Big Five personality questionnaires, as agreeableness reflects an interpersonal approach orientation coupled with a desire to attain closeness with others despite costs to the self, which amounts to a close approximation of trust (Costa and McCrae, 1992; Colquitt and Scott, 2007; Graziano and Tobin, 2017).

In large-scale surveys, self-report measures of trust are also commonly used, ranging from single-item measures that have been included in national (e.g., Gallup, Pew, and the General Social Survey) and international (e.g., European Social Survey, Pew Global, and Our World in Data) surveys for several decades to longer self-report measures for use in smaller studies (e.g., Rotter, 1967; Yamagishi and Yamagishi, 1994; Couch and Jones, 1997). The most widely used assessment of trust is the social trust question, originating from Rosenberg's (1957) classic Faith in People scale: “Generally speaking, would you say that most people can be trusted or that you can't be too careful in dealing with people?” To remedy the specificity problem, some researchers have added items asking about the trustworthiness of neighbors, strangers, friends, and other types of individuals directly (Helliwell and Wang, 2011).

In sum, trust has been adequately assessed in behavioral observations, psychometric scales, and single items in large surveys. However, experimental and survey research, when adequately powered, are costly to conduct, in addition to being potentially obtrusive. Recognizing the usefulness of measuring trust but being cognizant of barriers to large-scale measurement, researchers have urged for better assessment methods (Nannestad, 2008; Dinesen and Bekkers, 2017).

The advent of large online repositories of linguistic data and machine learning methods are now beginning to make measuring trust unobtrusively at scale possible. Automatic language-based assessments using social media data have provided valuable insights into people's personalities, emotions, and behaviors (Kern et al., 2014; Park et al., 2015; Curtis et al., 2018); identity (Rogers and Jones, 2021); mental health (Preotiuc-Pietro et al., 2015a); wellbeing (Schwartz et al., 2013a, 2016); and physical health (Eichstaedt et al., 2015). On social media, people frequently report their in-the-moment attitudes, feelings, and thoughts about the events that occur in their lives. Once a language model has been trained on scores for a given construct, linguistic patterns can be used to predict scores on that construct over other users using language alone (Schwartz et al., 2013b; Park et al., 2015; Preotiuc-Pietro et al., 2015b). We follow this approach to measure trust based on language shared on Facebook.

The current study examines the possibility of measuring trust at scale by developing and validating a language-based assessment of generalized trust. This is done in five stages. First, we derive a measure of self-reported trust. Second, we explore the language of trust by looking at how language varies across trusting and distrusting individuals, as well as variation across age and gender. Third, we build and validate a language-based machine learning model of trust. Fourth, this model is then applied to ~6 million geotagged Twitter users, providing a large-scale assessment of trust at the county level (i.e., regions within the United States), which we then validate. Finally, we use the county-level trust measure to gain new insights into community-level health, wellbeing, and political ideology.

The analytic sample was drawn from 154,000 users who directly consented to take a personality survey and share their Facebook language data (Kosinski et al., 2015). After consenting, participants completed a Big-Five personality questionnaire based on the International Personality Item Pool proxy for the NEO Personality Inventory (NEO-PI-R; Costa and McCrae, 2008), ranging from 20 to 336 items. Participants could choose to share their own Facebook statuses across the past 4 years for research purposes. Facebook statuses are short autobiographical posts that contain salient personal or emotional information, customarily shared only with individuals designated as “friends.” The use of the de-identified data set, including consenting users only (not their friends), was approved by the University of Pennsylvania's institutional review board (protocol #813820). The institutional review board deemed this low risk and granted exemption status, including no special requirements for adolescents beyond the same assent that all participants gave. The current study included a subset of users who (a) indicated English as a primary language, (b) completed the 100-item version of the International Personality Item Pool (IPIP) survey (which includes trust-relevant items), (c) indicated their age and gender, and (d) had at least 1,000 words across all of their status updates, as this threshold is common in order to have an adequate sample of each participant's language (Schwartz et al., 2013b; Kern et al., 2016). The resulting sample included 16,487 users (56.8% female; mean age 23.3 years, SD = 9.4 years, range = 13–80) who had shared 3,300,928 Facebook status updates (median status updates = 150) from January 2009 to November 2011.

We gathered publicly accessible data on geographic regions via Twitter and from national surveys (e.g., CDC and Gallup).

The self-reported trust measure is derived from three trust items in the 100-item version of the IPIP personality questionnaire (as described earlier). We assess convergent validity by comparing our measure to other trust measures: the Yamagishi inventory and the NORC-GSS inventory. Next, we correlate the trust measure with external criteria: Big 5 personality, wellbeing and health measures, self-monitoring, and impulsivity. Finally, we explore how trust is related to social media activity, such as likes, friend network size, and tags in photos. All external criteria are described in Section 2.2.2.

We used both top-down and bottom-up computational linguistic analysis approaches to identify linguistic correlates of trust. A top-down approach uses an established dictionary, or set of words that reflect categories, which are developed a priori based on theory. The most commonly used set of dictionaries in psychological science comes from the Linguistic Inquiry Word Count (LIWC) program (Pennebaker et al., 2015). The LIWC is a text analysis application developed to capture multiple psychological dimensions by counting the frequencies of words in a wide variety of categories. These frequencies can then be correlated with other constructs or used as predictors (Pennebaker and King, 1999). We used the LIWC 2015 dictionaries and considered correlations between trust and each category's frequency.

We also used a bottom-up approach—differential language analysis (DLA; Schwartz et al., 2013b)—which identifies language that characterizes generalized trust through semantically similar language clusters. DLA consists of three main steps: linguistic feature extraction, correlational analysis, and visualization, which are described in the following subsections (see Schwartz et al., 2013b; Kern et al., 2016, for further details). Both the top-down and bottom-up approaches use the open-source Differential Language Analysis ToolKit (Schwartz et al., 2017) for linguistic feature extraction, correlation analysis, and visualization.

To develop a language-based assessment of trust, we used techniques from statistical learning theory based on penalized (ridge) regression (Hastie et al., 2009). Following the well-validated approach of Park et al. (2015), we use three types of information as features (i.e., independent variables) for the statistical model: (1) relative frequencies of the 1- to 3-word phrases, (2) binary-transformed 1- to 3-grams (1 meaning the user used the word/phrase at least once, zero otherwise), and (3) relative frequencies of topics. We extracted these features from the 19,445 users that had three-item trust scores. We note that the features are extracted across the lifetime of posts across each user and, thus, assess trait-level trust for each person.

We used the same feature selection steps as Park et al. (2015), which consist of (1) selecting features with at least a small univariate correlation with the trust score (based only on the training data) and (2) performing a principal components analysis on the remaining features and limiting dimensions to 10% of the number of observations. These steps were performed independently for each of the three types of features (1- to 3-gram frequencies, binary 1- to 3-grams, and topics).

We established model accuracy by evaluating it over a hold-out set of N = 438 users that took the longer, 10-item version of trust (a so-called test set; Zamani et al., 2018). The “training set” used to fit the model consisted of the other N = 19,445 users. We use the larger three-item trust data for training, as it is often worthwhile to accept some error in the training set (i.e., due to answering fewer trust items) for a larger sample of observations (Schwartz and Ungar, 2015; Kern et al., 2016), while the more reliable less erroneous 10-item trust is used to establish the accuracy of the model. Accuracy is reported as the mean squared error (MSE) and the disattenuated product-moment correlation coefficient between our model's predictions of trust score and the trust score according to the 10-item questionnaire over the 438 held-out test-set users.

To consider test–retest reliability, we separated time into four periods (time 1: July–December 2009; time 2: January–June 2010; time 3: July–December 2010; time 4: January–June 2011), with an equal number of users (n per bin = 2,370).

Finally, we consider convergent and divergent validity by examining the relationship between our trust measure and the facets of agreeableness: trust, morality, altruism, cooperation, modesty, and sympathy. This was done in a sample of 414 individuals who completed the 336-item IPIP.

After establishing the validity of the language-based assessment of trust at the individual level, we applied that model to domains and data sets for which collecting generalized trust measures with traditional methods is infeasible, in this case feature sets extracted from location-mapped Twitter data at the U.S. county level (the CTLB, described earlier). This data set contains county-level language features, identical to the features described in Section 3.3: 1- to 3-gram words and phrases, both relative frequencies and binarized, and normalized topic usage (using the same LDA topic model). The language-based trust model described earlier (i.e., a trained regression model) can then be applied to the county-level language features, producing a trust score for each county.

To validate these county trust estimates, we compared our county-level language estimates against the Gallup–Sharecare Wellbeing Index. Items were chosen such that we predicted a positive relationship with trust.

We then correlated the language model-based estimate of community-level generalized trust with other community-level outcomes from large-scale census/survey, demographic, and environmental data. We use the same correlation method as at the individual level: a linear regression between estimated generalized trust as the independent variable and the county-level outcomes as dependent variables. The magnitude of standardized beta represents the effect size of the relationship, and their corresponding p-values are corrected for multiple comparisons using the FDR procedure (Benjamini and Hochberg, 1995). As defined previously, these outcomes include sociodemographic information (income, education, etc.), health (percent obese, mortality rates, etc.), and voting trends (percentage voting Republican in the 2016 presidential election as compared to 2012, voter turnout, etc).

The trust measure demonstrated adequate reliability (α = 0.73) and was strongly correlated with other trust measures (Table 1), such as the Yamagishi inventory (r = 0.77) and the NORC-GSS inventory (r = 0.70; correlations with individual items in Supplementary Table S1, including the trust item from Rosenberg's Faith in People Scale) in a separate sample (N = 1,041, mean age = 33.3, 55% female).

For the Facebook sample (N = 16,487), Table 2 provides descriptive information for the three items and the other variables available in the data set and correlations between these measures and survey-based, as well as language-based, trust. Here again, we observed strong convergent validity. As expected, trust is positively associated with related constructs such as fair-mindedness and self-disclosure. Among personality traits, the largest association is with overall agreeableness (agreeableness with the trust-facet items removed), although this correlation also demonstrates substantial independence between agreeableness and trust. Emotional stability was also strongly related to trust. Individual-level trust showed expected associations with health (physician visits, days sick) and wellbeing (life satisfaction). Our data also allowed for the exploration of some weak yet interesting associations with social behavior on the platform.

To assess the model accuracy, we compare the language-based trust estimates to self-reported trust in the test set. Results show that the model is accurate at MSE = 0.66 and product-moment correlation =0.49 (0.42, 0.56). These results are in line with accuracies predicting psychological constructs from social media text. Park et al. (2015) showed correlations between 0.35 and 0.43 between language-based personality estimates and self-reported personality.

For test–retest reliability, results are shown in Table 4 and were comparable to longitudinal correlations for other personality characteristics (Park et al., 2015). Scores were strongly correlated across the time points (range = 0.57–0.74), with decreasing strength over time.

In terms of convergent and divergent validity (i.e., comparing our trust measures to facets of agreeableness), the language-based measure was very strongly correlated with trust (β_survey = 0.77) and moderately correlated with overall Agreeableness (β_survey = 0.53), with the strongest correlation with the trust facet, and the weakest correlation with modesty (β_survey = 0.12) and mortality (β_survey = 0.12; see Table 5).

We applied the Facebook-derived model to geotagged Tweets and mapped them to U.S. counties (see Figure 6), with higher trust in blue and lower trust in red. Strong regional variations emerge (see Supplementary Figure S5 for a map in which sex, age, race, population, log density, and log income are controlled for).

We compared our county-level language estimates against the Gallup–Sharecare Wellbeing Index. Items were chosen such that we predicted a positive relationship with trust. Table 6 shows the items with their predicted and actual relationships with trust. All but two items (Someone in your life encourages you to be healthy and Hours spent socially) matched our predictions.

Further county-level analyses (see Table 7) revealed that trust was associated with more education (greater rate of high school graduation), higher income, and greater population density. More trusting counties had a greater percentage of individuals living in committed, stable relationships, as indexed by higher marriage and same-sex household rates, while less trusting counties had a greater percentage of separated individuals. This mirrors findings at the individual level (Helliwell and Wang, 2011). Areas with more Evangelical Protestants were lower in trust, while areas with more mainline Protestants and Catholics were more trusting, again confirming previous findings at the individual level (see Supplementary Table S3; Welch et al., 2004). Counties with more income inequality showed lower trust, in agreement with previous findings (Kawachi et al., 1997). These results suggest that trusting communities are generally more affluent, educated, and religious. Importantly, trust is low in counties with ethnic, economic, and political diversity.

Strong links to health and wellbeing were evident as well. Trusting counties had better self-rated mental and physical health, confirming what has previously been observed at the state level (Abbott and Freeth, 2008). Trust was also associated with healthier lifestyles (including obesity and percentage of smokers) and greater longevity. Unlike individual-level research that found social capital to be protective against excessive drinking (Takakura, 2011), counties with greater rates of excessive drinking tend to be higher in trust. Counties with greater trust also enjoy greater life satisfaction (congruent with findings at the individual level; Helliwell and Wang, 2011).

Finally, we examined county trust in relation to county political factors (see Table 8). Trusting counties had higher voter turnout in the 2012 and 2016 presidential elections. Distrusting counties had a higher Trump vote gain vs. both the Romney vote (2012) and the average Republican vote in the previous four elections (2000, 2004, 2008, and 2012). Donation partisanship (more donations associated with a single party) was associated with more trust.

Consistent with prior research, age and gender correlations with trust were relatively weak (Feingold, 1994; Helliwell and Wang, 2011). The positive association found with fair-mindedness is congruent with the previously suggested association between trust in others and one's own trustworthiness (Ben-Ner and Putterman, 2001; Hardin, 2004). Consistent with prior research, those higher in trust were higher in subjective wellbeing (Helliwell and Wang, 2011) and reported fewer sick days and physician visits, suggesting that more trusting individuals also may be healthier.

We also explored the association of trust with Facebook behaviors. Users may “tag” other users in posts being shared (including images and posts referencing social events). We found that more trusting individuals “tagged” more people and had more “friends,” perhaps suggesting that trusting users are comfortable sharing personal content with more people.

Associations with LIWC 2015 categories provide a parsimonious overview of the language of trusting individuals, suggesting that those higher in trust express more positive emotions and reference more interpersonal affiliations. They seem to be both engaged in their social and professional lives, referencing not only work life as well as achievement but also relaxation, home, and family time. LIWC results are in line with trusting individuals leading relatively balanced lives. Those lower in trust express anger and negativity, swear, and discuss risk. Of particular note, greater generalized trust was not associated strongly with greater use of social words, in contrast with previous work suggesting trusting individuals are more social (Hardin, 2002). However, upon closer inspection, it turned out that the lack of association with social terms was driven by the negative effect of othering terminology (they and people were associated with distrust) that countered positive associations with other social words (e.g., we). This motivates the idea of observing open-vocabulary analyses that do not make a priori assumptions about words and group membership (Schwartz et al., 2013b).

Open-vocabulary approaches yielded more specific language associations. We found that trusting individuals mentioned travel (flight, packing, and airport), supporting recent evidence that a breadth of travel (number of times traveled) is associated with generalized trust (Cao et al., 2014). Trusting individuals appeared integrated into cohesive social networks due to their more frequent mention of them (family, kids, and couples), which seemed to involve initiating or participating in social outings (free, pm, interested, and march) and taking an optimistic, happy, and future-oriented outlook on life (excited and tomorrow) when approaching others. Trusting individuals appeared to be more socially engaged (meeting, event, board, and council) and prosocial (amazing, great, and cheers), perhaps taking an active role in making a positive difference for their social networks, communities, or causes they care about, for example, being a council member and attending meetings. This suggests the benefits of trust not just for the individual but also for employers and the greater community through positive effects on health, wellbeing, and overall flourishing.

The language of distrusting individuals was different, marked by misanthropy (stupid, annoyed, people, and bitches), hatred (hate, hating, and hatin), and swearing (fuck, damn, and ass) about or at others, but it was also indicative of social isolation, emotional pain, and sadness (pain, broken, inside, and empty), suggesting that these negative views of others are deeply upsetting and alienating. It is possible that those who experience rejection or alienation from others (e.g., bullying, discrimination, and rejection) develop lower levels of generalized trust. As to be expected, individuals low in trust also appeared to be preoccupied with others' flaws, such as lying and selfishness (eyes, truth, and lies), and to frequently experience conflict and discord in their relationships with others (people, don't care, and rude). A focus on the self is also characteristic of those low in trust (I and I'm).

Distrusting adolescents mentioned loneliness, disappointments, and feeling misunderstood, as well as thoughts about pain and death. During early adulthood, for those low in trust, pain topics still prevailed but were joined by expressions of not caring (anymore) and some aggression, while post-college aggressive sentiment dominates along with swearing. From adolescence into young adulthood, distrust appeared to shift from being more self-focused to being more other-focused, suggesting that individuals low in trust might perceive themselves as the cause for their discord with others and then blame others and become disillusioned later in life. These patterns align with theory and research suggesting that the younger years are more important for the formation of trust (Erikson, 1963; Inglehart, 1997; Putnam, 2000) before pain and disappointment morph into disengagement and anger. Those lower in trust also mentioned the words people and they more, which tends to be mentioned more in negative contexts. This holds for distrusting males and females (Figure 4) and all age groups (Figure 5). Following the construal level theory of psychological distance (Trope and Liberman, 2010), it is possible that those lower in trust think more often about people as abstract and larger collectives, while those higher in trust are more prone to focusing on individuals.

Future research could directly explore the malleable aspects of generalized trust (Glanville and Paxton, 2007). It would be interesting to examine changes in the level of trust within individuals over time, both as a function of key life events and as impacted by changes in the individual's social network. Whereas such an analysis would be almost impossible with self-reported information, requiring an intensive longitudinal design, the temporal and network-based structure of social media can allow such a dynamic analysis. Using multilevel modeling approaches, it would be possible to identify the factors that lead to the rise or demise of the social trust of individuals and communities, individual–community fit based on trust, and the success of interventions aimed at increasing social trust. A deeper understanding of how social trust is negotiated, maintained, reduced, and increased as a function of time will be crucial for determining how social trust can be improved both for individuals and for communities.

County-level correlations (see Table 7) showed that more affluent areas and those with greater income equality are higher in trust, consistent with prior research (Nannestad, 2008). Trusting communities were also generally healthier and happier than distrusting communities, suggesting the importance of social cohesion and cooperation on regional health. A culture of mutual trust and respect may affect the health of its residents in a myriad of ways, from relying on health care and increasing one's family income to visiting doctors and following doctor's orders to taking advantage of public parks and other spaces to socialize and exercise. It has been suggested that building communities that are safe, walkable, and encourage positive social interactions may benefit both the sense of trust and the economic, physical, and mental wellbeing of the community (Jackson, 2003; Mason, 2010).

A notable exception of these positive effects of generalized trust at the community level is seen in a positive association with excessive drinking, a generally socially negative health behavior (Kuntsche et al., 2017). While it is possible that some third variable explains this relationship [although we controlled for socioeconomic status (SES) and region], it is plausible that trust may make people more comfortable to drink, as they are less concerned about interpersonal risks or that trusting communities celebrate together more. The role of alcohol in forming social bonds and trust is worth further study.

Otherwise, areas with greater distrust were those at greater risk of being unhealthy and premature death. It might be that these communities are marked by infrastructural deficits that do not allow for easy access to social interaction (e.g., long distances to recreational centers for social activity and exercise coupled with poor walkability and access to public transportation, a lack of social events, and a lack of safe, public places for gathering; Cohen and Inagami, 2005; Diez Roux and Mair, 2010) or that a culture of negativity makes it unattractive to spend time being active and interacting with others (Eichstaedt et al., 2015). This would be in line with suggestions that designing public spaces to make walking and socializing with others attractive will encourage embeddedness in social groups and promote health and longevity (Buettner, 2008).

Finally, Table 7 shows the relationship between the COI and the county-level outcomes. This index (and similar disadvantage indices) is currently used to allocate resources and guide policy decisions (Lou et al., 2023). The effect sizes for the COI are slightly larger, which is to be expected because this is a composite index built from 29 U.S. Census variables, as opposed to a language-based psychological construct. Nevertheless, the magnitude and direction effect sizes for the language-based trust measure and the COI are comparable, showing that our trust measure could be used in similar real-world settings to help develop effective interventions or even guide policy decisions. For example, the language of highly trusting individuals contains references to traveling, community events, and time with friends and family. This may suggest that for trust to be built, individuals may need to connect both with their own community more as well as travel outside of it, although more research is needed to draw firm conclusions on this point. Furthermore, continuous monitoring through language assessment could be deployed in aid to those who wish to reverse the general decrease in trust in the United States. Best practices for fostering trust across social, economic, and political divisions need to be established, shared, and tested. While causality has not been addressed here, the correlations suggest that fostering trust may well be important for improving the health and prosperity of communities. This may be particularly the case in highly fractionalized communities.

In the United States, confidence in institutions has been on the decline over the past several decades (Gallup, 2022). One way to examine the implications of this trend is to consider the relationship between trust and voter preference. We found the relationship between trust and party preference shifted between the 2012 and the 2016 presidential elections. The correlation between trust and Republican vote share moved away from the positive toward the negative. More aligned were the county trust levels and the votes gained by the 2016 Republican candidate (Donald Trump) over previous Republican candidates. Trump was supported more in counties that used language indicative of high distrust.

Donald Trump's election win was surprising to many (Flegenheimer and Barbaro, 2016). Early work to explain the result employed postelection surveys and focused on demographic segments (especially the “white working class”), ascribing motives based on personal identity (Morgan and Lee, 2018; Schaffner et al., 2018). The prospective data-driven approach employed here, however, enabled a window into trust based on verbal behavior during the campaigns. Together, the open-vocabulary method and large social media language data set covered orders of magnitude more people and more open-ended themes for hypothesis testing. Importantly, the data are longitudinal, providing data before, during, and after the election.

Two possible hypotheses follow from the pattern of results. First, communities lower in generalized trust prefer populist candidates (those that attempt to appeal directly to voters and circumvent parties, the media, and other institutions) to establishment candidates. Second, decreasing levels of generalized trust lead to lower vote shares for incumbent and incumbent-party candidates. These predictions are speculative—based on one surprising election result, but the real-time and local nature of the social media data could allow for future evaluations to be tested in a principled way—by predicting which type of candidate will be unexpectedly popular both locally and nationally.

Several limitations of this study should be acknowledged. First, at the time of data collection, social media users tended to be younger than the general population (Duggan and Smith, 2013), and, accordingly, most participants in our sample were young adults (median age = 20, range = 13–80). Still, we are able to control for age in our analyses, and due to the overall size of our data, we still have more participants 30 and older (N = 3,354) than most samples in trust research, which primarily occurs with university students.

Second, the current article is based on a survey-based measure of trust that utilizes trust items taken from the NEO-PI-R (Costa and McCrae, 1992). This scale is one of the most established measures of trust and has been used in previous research (Colquitt and Scott, 2007; Graziano and Tobin, 2017) and was highly correlated with other established trust scales. However, it is possible that if we fit our language-based model to a different scale, then it would have slightly different properties. Trust research has been criticized in general for its flawed assessment (e.g., Glaeser et al., 2000; Evans and Revelle, 2008; Nannestad, 2008), and a social media approach provides a valid alternative for future research that reduces problems with traditional self-report survey measures, including recency effects, recall bias, difficulty accessing diverse populations, and cost (Park et al., 2015; Vazire, 2015).

Third, a key question and potential limitation for all online behavior-based research is whether it reflects their offline personality and whether behavior online is consistent with offline behavior. Research establishing the validity of social media methods has demonstrated that there is a good indication that users tend to present themselves accurately to their social network (Kosinski et al., 2013).

Fourth, we identified a variety of correlates at the individual and community levels. While the multi-test corrections to significance give confidence to the reliability of these results, they do not suggest causal relationships. Other studies, using a variety of methodologies and samples, would replicate and refine the pattern of results revealed through our analyses using more targeted hypotheses.