deltie.net archive

Extracted abstract

Trust is a foundational concept of contemporary sociological theory. Still, empirical research on trust relies on a relatively small set of measures. These are increasingly debated, potentially undermining large swathes of empirical evidence. Drawing on a combination of open-ended probing data, supervised machine learning, and a U.S. representative quota sample, our study compares the validity of standard measures of generalized social trust with more recent, situation-specific measures of trust. We find that survey measures that refer to "strangers" in their question wording best reflect the concept of generalized trust, also known as trust in unknown others. While

Introduction

Generalized social trust is one of the fundamental concepts in contemporary social theory (Coleman 1994; Herreros 2004; Putnam, Leonardi, and Nanetti 1994; Schilke, Reimann, and Cook 2021; Smith 2010; Sztompka 1999; Uslaner 2002) and scholarly interest in this concept has grown alongside the increasing number of studies on social capital and social cohesion, as trust is considered a main indicator of these concepts (Larsen 2013; Portes and Vickstrom 2011; Van Deth 2003) . Consequently, empirical research investigating the causes and consequences of trust has multiplied (Buskens and Weesie 2000; Cook and Cooper 2003; Dinesen 2012; Dinesen, Sonne Nørgaard, and Klemmensen 2013; Dinesen and Sønderskov 2015; Sønderskov 2011) . At the same time, the underlying empirical research program relies on a relatively small set of established survey measures, some of which date back to the 1940s. In recent years, we have seen a growing debate about the validity of these measures, particularly regarding their ability to capture the same concept across all individuals (Bauer and Freitag 2018; Delhey and Newton 2005; Delhey, Newton, and Welzel 2011; Ermisch et al. 2009; Nannestad 2008; Robbins 2019; Sturgis and Smith 2010; Torpe and Lolle 2011) .

Our study aims to address this debate by investigating the validity of survey measures of generalized social trust. In doing so, we make several contributions to current research. First, we evaluate three classic trust measures in a U.S. sample, thus extending previous work that examined fewer measures using data from the United Kingdom (Sturgis, Brunton-Smith, and Jackson 2019; Sturgis and Smith 2010) . All three measures have been used to measure generalized social trust, specifically trust in unknown others (Sønderskov 2011; Uslaner 2002) . The first measure is known as the "most people question" (Rosenberg, 1956) , which poses the query "Generally speaking, would you say that most people can be trusted, or that you can't be too careful in dealing with people?". The second measure, referred to as the "people first time question" (e.g., Torpe and Lolle 2011) , asks respondents about their level of trust in people they meet for the first time. Both of these measures have been established and utilized in numerous large-scale surveys. In contrast, what we call the "stranger question" (Robbins 2019 (Robbins , 2021)) , which is "Imagine meeting a total stranger for the first time. Please identify how much you would trust this stranger" is a more recent alternative and hopeful contender, expected to alleviate some of the problems that appear to characterize the former two. Our study revolves around exploring the validity of these three measures and scrutinizing whether they genuinely measure trust in unknown others, thus identifying possible measurement errors that might influence estimates of trust levels. To achieve this, we designed a survey experiment in which the different measures were randomly assigned to respondents. Our main findings are derived from using open-ended questions that ask about respondents' frames of reference, what we call associations, underlying their response.

Second, we contrast classic measures of generalized social trust with situative measures of trust. Such measures differ from the classical ones in that they specify a more refined trustee category (e.g., "most people" is replaced with "stranger") as well as some behavior at which the expectation is directed (e.g., "keeping a secret"). Ideally, such measures are able to provide a higher degree of interpersonal comparability since they leave less room for different interpretations by the survey respondents. We are the first to probe such measures and provide evidence on whether validity and comparability increases when these measures are used.

Third, we explore the sentiment of associations, a dimension that has been neglected so far in trust research. Theory assumes that trust in known others is higher due to effects of in-group bias and reciprocity (Vollan 2011) , which is supported by empirical evidence (e.g., Bauer and Freitag 2018; Sturgis and Smith 2010) . However, independently of whether respondents refer to known or unknown others, associations may also vary in terms of their sentiment, for example whether they are positive or negative.

Fourth, we extend the methodological toolbox that is used to evaluate the validity of survey measures, using a combination of open-ended probing questions (e.g., Behr et al. 2012 Behr et al. , 2017;; Meitinger and Kunz 2022; Neuert, Meitinger, and Behr 2021) and automated text analysis (e.g., Schonlau and Couper 2016) . The data we labeled and the resulting supervised classifiers we built are suitable for future applications.

Theory, Hypotheses, and Previous Research

Associations With Known and Unknown Others

Generalized social trust is often referred to as trust in the generalized other and can be described as trust in individuals who are unfamiliar or unknown (Sønderskov 2011; Stolle 2015; Sturgis and Smith 2010; Uslaner 2002:52) . Stolle (2015) for example emphasizes the need to distinguish the scope of generalized trust from trust toward people one personally knows (Stolle 2015:398) . Notably, other accounts have chosen to expand the concept of generalized or social trust trust to encompass a wider range of trustees, such as trust "in people in general" (Yamagishi and Yamagishi 1994:146) , or as trust in the "average person [one] meets" (Coleman 1994:104) . Our study, however, uses the understanding of generalized trust that stresses the difference between generalized and particularized trust. Particularized trust is defined as "[…] trust found in close social proximity and extended toward people the individual knows from everyday interactions" (Freitag and Traunmüller 2009:784) , including family members, friends, neighbors and co-workers (Freitag and Traunmüller 2009:784) (i.e., known others), whereas generalized trust encompasses "[…] those beyond immediate familiarity, including strangers" (Freitag and Traunmüller 2009:784) (i.e., unknown others) . In this study, we argue that when conceptualizing generalized trust, it should ideally be measured as trust towards unknown others.

Currently, the measurement of trust primarily relies on survey questions, although behavioral measures and their combination with survey measures have gained popularity (Barr 2003; Ermisch et al. 2009; Ermisch and Gambetta 2010; Fehr et al. 2002; Naef and Schupp 2009) . Various different questions are used in different large-scale surveys. Undoubtedly, the standard measure is the so-called "most people question" which inquires whether most people can be trusted. Different versions of this question were used in thousands of influential studies and underlying surveys, such as the General Social Survey, the World Values Survey or the European Social Survey.

However, the measurement of trust using the most people question has been subject of many debates (cf. Bauer and Freitag 2018) regarding various aspects, such as scale length or balance (Lundmark, Gilljam, and Dahlberg 2016) , and the frames of reference employed by respondents when answering it (Delhey, Newton, and Welzel 2014; Nannestad 2008; Sturgis and Smith 2010) . These frames of reference, what we call associations, are important as they are linked to the conceptual validity of a measure. Conceptual validity increases when the respective survey questions capture generalized trust without specification or measurement error. Figure 1 depicts our main argument regarding these associations.

When employing trustee categories such as "most people" in standard trust measures, it is probable that distinct associations may arise among different respondents. For instance, in the illustrated example presented in Figure 1 , respondent Hanna envisions a friend, while Hans envisions a stranger when answering the corresponding survey question. This scenario highlights the ongoing debate on equivalence and whether the concepts in the questions are uniformly interpreted by all respondents (Bauer and Freitag 2018) . Consequently, due to these varying associations, Hanna's response reflects particularized trust, resulting in a specification error, while Hans's response more closely aligns with the notion of the generalized other. These differences in associations can lead to divergent responses on the trust scale between two individuals (e.g., Hans and Hanna) or even within the same individual at different points in time (depicted by the dashed line in Figure 1 ).

Given that the conceptual definition of generalized (and particularized) trust refers to the distinction between known and unknown others, our study aims to identify the associations arising from the specific wording of survey questions. Empirical evidence in that direction is given by Sturgis and Smith (2010) . In examining the most people question using think-aloud probing, they describe six higher-order topics they found respondents to associate with the term "most people." The two largest categories they found by manually classifying responses to their probing question were "known others" (42 percent) and "unknown others" (22 percent). 1 In a similar approach, Bauer and Freitag (2018) surveys student samples from Switzerland using a probe that asks respondents who they had in mind when answering the most people question. The open-ended text answers reveal that "respondents do not necessarily tend to think of strangers or people that are unknown to them. Many think of situations (e.g., meeting someone in the train/street) or of people they know (e.g., friends, family members, etc.)" (Bauer and Freitag 2018:9) . Lastly, Uslaner (2002:72-4) , as part of the 2000 ANES Pilot Survey, investigated the most people question via think-aloud techniques and showed that 58 percent of the respondents referred to a "general worldview" while 23 percent mentioned "personal experiences." While personal experiences do not necessarily involve known others, the 2002 ANES data was also coded into more fine-grained categories by Johnson (cf. ANES 2000): 8 percent of respondents referred to family members, 11 percent to co-workers, and 12 percent to neighbors.

The present study compares three established measures of generalized social trust, the "most people question" (M1), the "people first time question" (M2), and the "stranger question" (M3). Next to M1, M2 is the second most common generalized trust measure used in many large-scale surveys, such as the World Values Survey or the Socio-Economic Panel in Germany. M3 is a more recent measurement approach, which is not yet part of larger surveys, and was developed with the aim that respondents imagine strangers in their answer (Robbins 2019 (Robbins , 2021)) . Our particular interest for each of these measures lies in the proportion of respondents who think of personally known others (short: known others), when answering expressed as

p k = 1 n n i=1 Y i ,

where Y i is a dummy that indicates whether individual i thought of known others (1) or unknown others (0) in their response. Importantly, across the three measures M1-M3, the trustee category is gradually refined. M1 is fairly vague and only refers to most people. M2 already specifies that respondents should think of first-time encounters. M3 further specifies the trustee category by clarifying that the trustee category encompasses strangers. We expect that explicitly referring to "people you meet for the first time" (M2) or "a total stranger you meet for the first time" (M3) as compared to "most people" (M1) may increase the proportion of respondents thinking of others they do not know (1p k ). Furthermore, we expect that using the stranger-wording (M3) should increase this share even more than using the people-wording (M2). In our view, the people-wording is more likely to produce associations of situations where the respondent has had first-time encounters with persons that are well-known by now. For instance, respondents may think of a first-time encounter with friends, work colleagues or relatives or first-time encounters with persons who are already connected (e.g., first time meeting the new partner of a sibling). In contrast, the strangerwording should make it more likely that respondents think about situations in which they really don't have (or haven't had) any information about the trustee (e.g., encounters in the street). Eventually, we hypothesize that a refinement of the trustee category (most people → people you meet for the first time → a total stranger you meet for the first time), decreases the proportion of respondents in whom the association with known people (p k ) is evoked (H 1 ). Evidence for H 1 would be provided by statistically significant differences between those proportions:

p k,M1 > p k,M2 ; p k,M1 > p k,M3 ; p k,M2 > p k,M3 .

Additionally, following Sturgis and Smith (2010) , we also expect that individual associations with known others positively influence trust scores (H 2 ) across all three measures. For instance, when calculating the aggregate mean level of trust, y = 1 n n i=1 y i , where y i is an individual i's reported trust score, we could expect a positive difference in trust between the subset of respondents who think of known others and respondents who think of unknown others. Estimating such differences could help us identify the measurement error that is included in common aggregate estimates of trust scores.

Negative Associations

While trust research regularly discusses the impact of experiences on trust (Brehm and Rahn 1997; Cao, Galinsky, and Maddux 2014; Dinesen 2010; Freitag and Traunmüller 2009; Glanville, Andersson, and Paxton 2013; Glanville and Paxton 2007; Uslaner 2002 ), studies about trust measurement have neglected this dimension. On average, trust in known others is higher (Bauer and Freitag 2018; Sturgis and Smith 2010; Vollan 2011 )-as is also evidenced by measures that directly gauge trust in family members, neighbors, etc. (Freitag and Traunmüller 2009; Nannestad 2008) . Theoretically, however, this does not always have to be the case. In fact, some of the more important betrayals of trust in our lives may happen through people we know. For instance, a close friend may spill our secrets or a family member may fail to return a loan. Referring to Figure 1 , Hans's response may be based on a negative association as opposed to Hanna's response. Put differently, we may collect negative (or positive) experiences with known others just as we may collect negative (or positive) experiences with unknown others, that is, strangers. Independently from whether a trustee is known or unknown, individual associations that emerge when answering survey questions may vary in terms of their sentiment. Hence, we also want to measure the proportion of respondents who have negative associations, expressed as p n = 1 n n i=1 Y i , where Y i is a dummy that indicates whether individual i's association can be classified as negative (1) or not (0). 2 Again, the share of negative associations may depend on the measure we use. Since M2 (in contrast to M1) explicitly asks respondents to think of firsttime encounters ("people you meet for the first time"), we expect that this question wording may evoke more negative associations than the most people question. This could be either because respondents remember past first-time interactions that turned out to be negative and/or because we are generally taught to be careful in first-time encounters. M3, then, explicitly specifies the trustee as a stranger. The term "stranger" has a rather negative connotation in English compared to the more neutral terms "people" or "person." "Stranger danger" describes the idea that all strangers can potentially be dangerous. In countries such as Great Britain, stranger-danger education often conducted by local police force has the objective to teach children to refuse offers from strangers (Moran et al. 1997:11) . Postulating H 1 , we assume that M2 and M3 result in higher conceptual validity (i.e., lower share of associations of known others) which is desirable. However, finding that M3 or M2 in comparison to M1 result in more negative sentiment would be undesirable as it could indicate that using concepts such as "stranger" in M3 affects respondents' mindset.

We hypothesize that changing trustee categories (most people → people you meet for the first time → a total stranger you meet for the first time) increases the proportion of respondents who have negative associations (p n ) (H 3 ). Again, evidence for H 3 would be provided by statistically significant differences between those proportions: p n,M1 > p n,M2 ; p n,M1 > p n,M3 ; p n,M2 > p n,M3 . We also expect that negative associations should negatively influence trust scores (H 4 ) across all three measures. Thus, when calculating the mean level of trust y = 1 n n i=1 y i , where y i is an individual i's trust score, we expect a negative difference between the subset of respondents who have negative associations and those who do not have negative associations with M1, M2, and M3.

Situative Trust Measures

Empirical operationalizations of generalized trust, for example, M1-M3, depict trust as a "one-part relationship, where neither B [the trustee] nor × [expected behavior] enters explicitly" (Nannestad 2008:415) . In contrast, conceptual work argues that trust is a three-part relationship, in which A (truster) trusts B (trustee) with respect to some behavior X (Cook, Hardin, and Levi 2005; Schilke, Reimann, and Cook 2021) . Ermisch et al. (2009) criticized common survey measures of generalized trust to be too generic since the "[…] answers do not reveal either the reference group or the types of action or the stakes that respondents have in mind when making such an assessment" (Ermisch et al. 2009:750) . Their notion of trust includes a situative character, because they describe a trust situation to be characterized by "trust that someone will do X" (Ermisch et al. 2009:751; Ermisch and Gambetta 2010:4) .

The measures we investigate (M4.1-4.4) follow this conceptual work and include the context in which a trust decision takes place. This context entails two components, the trustee category, and the trustee's expected behavior in a certain situation. Importantly, the decision to trust in situation A may not carry over to situation B (Ermisch and Gambetta 2010:4) even though both situations involve the same trustee. We argue that situative trust measures may be able to solve some of the problems that characterize the vaguer standard measures of generalized trust. Since the latter do not specify either of the two components of context, respondents may simply fill in such specifications themselves.

Our study investigates situative trust measures introduced by Robbins (2019, 2021). These novel measures are based on the stranger question (M3) because they specify the trustee to be a stranger (cf. M3) (see Buskens and Weesie 2000; Yamagishi and Yamagishi 1994; Yuki etal. 2005 for similar approaches). Further, they specify the expected behavior of the trustee, namely keeping a secret (M4.1), repaying a loan (M4.2), providing advice on managing money (M4.3), and looking after a child/family member/loved one (M4.4). Unlike the stranger question (M3) that allows for varying interpretations by respondents, these situative measures provide a more specific context, leaving less room for ambiguity. This avoids situations where different respondents envision different scenarios, potentially leading to varying trust values (cf. Figure 1 ). Analogous to H 1 , we hypothesize that by specifying the trustee as a total stranger, as opposed to most people or people you meet for the first time, the proportion of respondents associating trust with known people (p k ) will decrease (H 5 ). As these situative measures are relatively new, we do not have specific expectations regarding the negativity of associations they may evoke or how they compare to each other. It is plausible that questions concerning money lending or money advice could elicit negative associations or memories. The question is, however, whether they do so systematically. Therefore, the empirical insights we present below are exploratory in nature.

Data, Experimental Design, and Methods

Sample

Our target population are U.S. citizens. Data was collected using a two-stage non-probability sample recruited by Prolific, a participant recruitment and payment software to conduct online surveys and experiments (Palan and Schitter 2018) . First, respondents were identified to be eligible according to quotas on self-reported gender, age, and ethnicity in accordance with the U.S. Census Bureau population group estimates from 2015. 3 Second, out of 43,131 panelists that were considered eligible, we continued to collect data until our target and final sample size of n=1,500 was reached. Respondents who did not complete the questionnaire (n=87, i.e., overall response rate of 95 percent) were excluded and replaced with other panelists who would fit the quotas. Summary Statistics for all variables and their comparison to population estimates can be found in Online Appendix A.1. The survey was fielded between July 14, 2021, and July 21, 2021. For each completed survey, we paid a wage of 9.60 USD/hour on average while the mean duration was 6.8 minutes.

Experimental Design and Measures

Our questionnaire design is depicted in Table 1 . Respondents provided their data via an online self-administered survey (created using formR, cf. Arslan, Walther, and Tata 2020) . The survey started with information on its objective and a consent form. Subsequently, respondents received two blocks of questions. Block #1 included the standard generalized trust measures with respective probing questions and Block #2 included situative trust measures with respective probing questions. Since we wanted to avoid priming effects (meaning subsequent answers might be influenced by previous questions) we used an experimental design in which the order of questions is randomized. Specifically, the order of Blocks #1 and #2 as well as the question order within these blocks was randomized. This design allows us to conclude that the differences we find between the trust measures for the outcomes we examined (i.e., the proportion of associations that refer to known individuals or are negative) are actually due to the wording of the question and not to the order of the questions.

Furthermore, data collected with this questionnaire allows for within-and between-person comparisons for each variable because each respondent received all available trust questions in Blocks #1 and #2 in a randomized order. To allow further examination of the role of question order despite the introduction of random question order, we can consider two data subsets: Subset 1 only includes respondents' responses to the first trust question they received (ignoring the order of the blocks) and is called "first question only" below; Subset 2 includes respondents' responses to the first trust question from the first block only and is called "first question and first block only" below. While there might still be priming from the preceeding block for Subset 1, this possibility should be excluded for Subset 2.

Block #1: Generalized trust measures and probing questions. In Block #1, we assessed generalized trust using three established measures: trust towards "most people" (M1), "people you meet for the first time" (M2), and "a total stranger you meet for the first time" (M3). These measures had different response categories: 7-, 4-, and 4-point scales for M1, M2, and M3, respectively. To ensure comparability, we employed min-max normalization, which rescales the responses to a range between 0 and 1 while preserving the original distribution. We treat the resulting variable as continuous for all our analyses. 4 The specific phrasing as well as summary statistics of these questions can be found in Online Appendices A.1 and A.2. Directly after respondents answered these closed-ended questions, each was followed by an open-ended probing question using the following wording (exemplary for M1): "In answering the previous question, who came to your mind when you were thinking about 'most people'? Please describe." Our specific interest here is to elicit who respondents had in mind when they were exposed to the three different trustee categories. 5

Block #2: Situative trust measures and probing questions. Block #2 included four situative measures that represent the Imaginary Stranger Trust (IST) scale developed by Robbins (2019 Robbins ( , 2021 Robbins ( , 2022)) . These measures specify the trustee category as well as the content of the trust relationship, overall aiming to reduce the vagueness we argued to find for the standard generalized trust measures from Block #1. The four items elicit trust in a total stranger met for the first time to, 6 (1) "keep a secret that is damaging to your reputation" (M4.1), ( 2 ) "repay a loan of one thousand dollars" (M4.2), ( 3 ) "provide advice about how best to manage your money" (M4.3), and to (4) "look after a child, family member, or loved one while you are away" (M4.4). Each of these items was rated on a 4-point scale. We applied min-max normalization to rescale these items to a range between 0 and 1. Again, the question order was randomized. Analogous to Block #1, the situative measures were also probed using the following wording: "In answering the previous question, who came to your mind when you were thinking about 'a total stranger you meet for the first time'? Please describe." To avoid memory effects as well as errors due to response fatigue, we only probed the situative measures that were randomly assigned to come first and fourth.

Methods

Table 2 illustrates the structure of our data. Due to the intra-person design, there are multiple (i.e., seven) measures of trust (indicated by the column Measure) for each respondent alongside their respective trust score (column Trust). Overall, we collected open-ended responses using five open-ended probing questions and received 7,497 out of potentially 7,500 text answers (column Probing Answer). 7 Online Appendix A.3 provides a detailed description of the open-ended text answers. Table 2 also displays the results for our classification of the openended responses (columns Associations (known-unknown others) and Associations (sentiment)). Both approaches are described in detail below.

Both classifications (i.e., known-unknown and sentiment) were achieved using automated text analysis, which in survey data research has become a popular alternative to manual coding (Esuli and Sebastiani 2010; Giorgetti and Sebastiani 2003; Gweon and Schonlau 2023) . In particular, we pursued a supervised classification approach in which randomly sampled subsets of text answers were manually labeled and only the remainder were automatically classified using fine-tuned BERT models.

For the known-unknown classification, we manually labeled a sample of n = 1,000 text answers, while for the sentiment classification, we increased this number to n = 1,500. 8 Both samples were a random selection of text answers from the generalized trust measures (see Online Appendix A.5.2 for further details). Based on previous implementations in the literature, we argue that these sample sizes are sufficiently large. 9 Both manual classification tasks were achieved using a hand-crafted coding scheme. For both schemes, the main distinction lies between two categories. In the known-unknown classification, category 0 was assigned when respondents mentioned individuals or groups of individuals that can be identified as "unknown others" in their text answer. Importantly, our primary focus was on identifying respondents' personal unfamiliarity with these individuals or groups, and not on the specific characteristics of these individuals/groups. For example, an answer that describes personally unknown others that have rather specific characteristics (i.e., tourists in ID 3139 in Table 2 falls into category 0). 10 Code 1, on the other hand, subsumes all statements that made mentions of "others known" to the respondent. Survey answers that had no references to either known or unknown others (e.g., "just people as a whole") were coded as 0, and survey answers with mixed references to both known and unknown others (e.g., "People I may run into everyday") were coded as 1. To label sentiment, the main distinction lies between "negative sentiment" (code 1) and "neutral or positive sentiment" (code 0). Online Appendix A.4 provides an overview of the coding schemes with examples and descriptions of all available codes.

The manual classification was carried out by three independent coders. All three coders assigned codes to the same 1,000/1,500 text answers, and conflicts were resolved by finding consensus between the coders or using majority vote.

For the remainder of text answers (i.e., n = 6,500/6,000), we fine-tuned the weights of two bidirectional encoder representations from transformers (BERT) models (BERT base model uncased version), using the manually coded data (n = 1,000/1,500) as training data. BERT (Devlin et al. 2019 ) is an empirically powerful machine learning technique that can be used for various natural language processing tasks (Devlin et al. 2019:1) . BERT comes with two attributes that are of special importance here: first, it is able to model contextual representations by incorporating both the left and right context of a document (i.e., bidirectional). Second, BERT provides pre-trained vector representations for words by using a deep, pre-trained neural network. These so-called embeddings suggest a representation for each term based on its context by using information from the entire input sequence. For our data, this could mean, for example, that terms that appear in the (pre-trained) context of "family," for example, brother and sister, are likely to be predicted as "known other." Last but not least, by using BERT, we aim at addressing the class imbalance that is present in our sentiment data insofar as few respondents (8.7 percent) have negative associations. BERT achieves higher class-wise accuracy in the presence of class imbalance than other ngram-based machine learning techniques (Gweon and Schonlau 2023) , and is further demonstrated to remove the need to use data augmentation techniques to mitigate problems of imbalanced data (Madabushi, Kochkina, and Castelle 2020) . 11 Importantly, the imbalanced data structure and its consequences does not call into question the effects we found but may have resulted in their slight underestimation. Online Appendix A.5.2 shows our findings when using the manually classified data only.

A detailed evaluation of the two classifiers in terms of accuracy, precision, recall, and F1-score is shown in Table 3 .

Alternative approaches with which we classified our data (i.e., regular expressions and random forest) can be found in Online Appendix A.6.

Results

Trust Scores Across Standard and Situative Measures

We begin by assessing the variations in trust scores obtained from our seven trust measures across different sample specifications (Figure 2 ). Regardless of the subsample, there is a gradual decline in trust from Measure 1 (most people Within-subjects ANOVA reveals that the generalized trust scores differed statistically significantly for the same individual for the three question wordings (F(1.7, 2,505) = 129, p < 0.001). 12 Additionally, situative trust measures M4.1-4.4 consistently exhibit lower trust levels likely owing to their emphasis on trust decisions where the truster has a lot to lose. 13 It is crucial to note that Figure 2 provides a descriptive overview of the seven measures concerning their sample means. The observed differences may be influenced by various factors, such as question interpretation, demand effects, and scale effects. In our subsequent analysis, we focus on examining one specific factor: the associations formed by respondents when answering our trust survey questions.

Associations Across Standard and Situative Measures

We start by examining the known-unknown dimension. Figure 3 displays the share of respondents who described associations of either known or unknown others across our seven measures. 14 In line with our expectation (H 1 ), the share of respondents referring to a known other statistically significantly decreases for M3 (i.e., 13 percent) while shares for M1 and M2 are similar (31 percent and 30 percent, respectively). The share of respondents referring to a known other again increases for our situative measures M4.1-4.4, however, none of these differences are statistically significant. Nevertheless, it could indicate that referring to specific situations and behaviors in those survey questions could increase the number of respondents who think of known others. This is undesirable from a conceptual perspective.

With regards to the sentiment dimension, we expected to find different shares of negative sentiment for each question wording (see Figure 4 ). In line with our expectations (H 3 ), the share of negative associations is higher for M3 (i.e., 8.7 percent) compared to M2 (7 percent). Not in line with our hypothesis, the share for M1 is higher (10 percent). However, none of these differences are statistically significant. Moreover, the share of negative associations remains similarly low for the situative measures, which is in accordance with the findings for M3 since the situative measures also describe the trustee category to be a "stranger."

In sum, we find that, across all seven measures, there are respondents who have associations with known others as well as associations of negative sentiment. However, strong differences between measures in terms of associations can only be found for the known-unknown dimension. The sentiment dimension seems less relevant. The two classification dummies only correlate weakly (r(7, 490) = -0.08, p = <0.001).

Associations and Trust Scores

Above we demonstrated that there is variation in associations across individuals. Next, we examine whether different associations affect the measurement values. Figure 5 visualizes the coefficients for a series of regression models (see Online Appendix A.9 for detailed regression tables). We estimated five models for each of our seven trust measures which are indicated on the left side. Two models are bivariate and only include one of the association dummies (e.g., Models #1 and #2 in Figure 5 ). We subsequently add covariates to these bivariate regressions (e.g., Models #3 and #4 in Figure 5 ). 15 Finally, the fifth model includes both dummies in one model and adds covariates.

In accordance with our expectations (H 2 ), we observe that associations with known others have a positive effect on trust for all of our three generalized trust measures M1, M2, and M3 (β #1 = 0.064; β #6 = 0.037; and β #12 = 0.023, respectively). While this effect is especially pronounced for M1 and M2 in terms of effect size and statistical significance (p < .001), it becomes smaller and less robust for M3. This may be due to the fact that M3 evokes associations with known people in fewer respondents than M1 an M2 do (see Figure 3 ), thus resulting in a smaller sample size of that subgroup, increasing the uncertainty of the corresponding estimate. In addition, adding the sentiment dummy as a control variable in Models #5, #10, and #16 (see Figure 5 ) does not mitigate the effect of the known-unknown dummy on trust.

In line with our expectation (H 4 ), we find that negative associations have a negative effect on trust for all of our three generalized trust measures M1, M2, and M3 regardless of the control set specifications (β #2 = -0.041, p < 0.01; β #7 = -0.066, p < 0.001; and β #13 = -0.049, p = 0.059, respectively). While the different generalized trust measures are not affected differently, we suggest that the role of negative associations for trust measurement requires future research.

Also for the four situative measures, the effects are in line with H 2 . Associations with known people have a positive effect on, for example, M4.4, trusting someone to watched a loved one (β #36 = 0.053, p < 0.001), or on M4.2, that is, trusting someone to repay a loan (β #24 = 0.053, p < 0.001). For the situative measures, however, while consistent with H 4 , we find smaller and less robust effects for our dummy capturing negative associations.

In sum, for the generalized trust measures, we find statistically significant effects in our hypothesized directions, namely that associations with known others (in contrast to unknown others) influences trust scores positively and that negative sentiment (in contrast to neutral/positive sentiment) influences trust scores negatively. Especially the effect of the dummy capturing the known-unknown dimension is undesirable from a conceptual point and its effect varies across measures of generalized trust. We can conclude that estimates based on the three classic measures-M1, M2, or M3-overestimate trust scores because they do not measure generalized trust for a significant share of the respondents. Without these respondents, our estimated trust averages would differ (namely by the coefficients we depict in Figure 5 for the bivariate models). The bias is smallest for the stranger measure M3 and all four of the situative measures seem to be characterized by the same problem.

Discussion and Conclusion

Generalized social trust is a foundational concept in the social sciences. However, there have been doubts about the validity of commonly used measures (Delhey, Newton, and Welzel 2011; Ermisch et al. 2009; Nannestad 2008; Robbins 2019; Sturgis and Smith 2010) . In our study, we examined various trust survey measures in a U.S. sample and explored how respondents answered those questions. To eliminate interviewer effects, we used a web probing approach (Behr et al. 2012 (Behr et al. , 2017;; Meitinger and Kunz 2022) . Open-ended probing (Neuert, Meitinger, and Behr 2021 ) is still a novelty in trust research, and similar data has so far only been collected in intervieweradministered settings (Sturgis and Smith 2010; Uslaner 2002) . The data collected through open-ended probing was analyzed using a supervised machine learning approach. Our findings can be categorized into four key aspects. First, our study revealed significant variations in overall and intra-individual reported trust levels across different question formats, and the question employing the phrase "most people" yielded the highest average trust score (cf. Figure 2 ). This finding suggests that the different question formats should not be considered interchangeable measures of generalized trust. However, it is important to note that Figure 2 provides only a descriptive overview, and our subsequent analysis centered on exploring the associations formed by respondents while answering the trust survey questions.

Second, we delved into the associations respondents made when responding to the questions. We described generalized trust as trust in unknown others, and argued that it should ideally be measured accordingly. Remarkably, a notable proportion of respondents (ranging from 13 percent to 31 percent, cf. Figure 3 ) incorporated thoughts of known individuals in their responses while answering classic trust questions, which is in line with previous research (e.g., Sturgis and Smith 2010) . Hence, for this particular group of respondents, classic trust measures actually do seem to capture what is commonly known as particularized trust (cf. Freitag and Traunmüller 2009) . In other words, for these respondents, our measures suffer from construct invalidity. However, the proportion of mentions of known individuals in responses decreased for the "stranger" question (M3), suggesting a higher degree of construct validity for this measure (in line with Robbins 2019 Robbins , 2022)) . Interestingly, compared to M3, the situative measures (M4.1-4.4) showed an increase in respondents thinking about known individuals (but still considerably smaller than in M1 and M2) (cf. Figure 3 ), despite being instructed to consider the trustee as a stranger. This outcome may be attributed to respondents drawing upon their past experiences to contextualize and anchor the given situations.

Thirdly, we conducted an examination of the influence of associations on trust levels. If confirmed, this would imply that trust estimates produced by specific measures (e.g., the "most people" wording) could be biased, potentially leading to an overestimation of generalized trust in diverse populations. Indeed, we found that respondents who reported thinking about known others displayed higher levels of trust across all three generalized trust measures (cf. Figure 5 ). The effects were less robust for the stranger question (M3), which might be due to the smaller share of respondents having known others in mind when answering. This is a desirable feature of the latter measure. 16 Overall, this finding demonstrates that differences in trust between individuals and over time may not be solely reflective of variation in the substantive dimension of trust. Instead, they might be influenced by specification errors and differences in how respondents interpret the question due to inter-individual differences in frames of reference.

Fourth, we also explored a hitherto neglected dimension-the sentiment of association. We found a relatively low proportion of respondents reporting negative associations which remained consistent across measures (cf. Figure 4 ). Against our expectations, M3, the stranger-question (without situations) does not seem to evoke more negative associations than the most people and people first time question. While negative associations did influence trust scores negatively, the effect was not uniform across measures and models (cf. Figure 5 ). These findings offer encouraging insights into measurement, yet we call for further research to explore whether specific question formats trigger more emotional responses or negative memories. Our study yields several key findings that not only allow us to draw valuable conclusions but also pave the way for future research directions.

Firstly, among the trust questions we investigated, our various "stranger" questions (M3 and M4.1-4.4) demonstrated the highest level of construct validity, as evidenced by the lower share of respondents thinking of known individuals. However, from an empirical perspective, we may question how many trust situations actually take place among total strangers. For example, the four situations in our study are more likely to take place among individuals who have some knowledge about each other (e.g., acquaintances). Certainly it can be challenging to pinpoint situations that entirely lack associations to known others, but we think that further theoretical work is necessary to classify based on whether a trust measure primarily pertains to strangers or also encompasses acquaintances. 17 Secondly, researchers should carefully consider various factors when selecting measures for their studies, aligning with their specific definition of generalized trust. Our findings indicate that M3 best captures generalized trust when defined as trust towards unknown others (cf. Figure 3 ). However, for those interested in interpersonal comparability, situative measures like the IST scale offer a viable alternative, since they explicitly define the concrete situation in which trust has to be placed and thus leave less room for different interpretations. Nonetheless, they demand additional questionnaire space due to longer item descriptions. 18 Generally, future studies could make use of additional, situative measures by using vignette designs. The resulting data could be analyzed in such a way, that one caclulates the average trust across a set of situative trust measures, yielding a score of what we call cross-situational trust (Bauer and Freitag 2018; Robbins 2022) . 19 However, we would also like to emphasize that the use of traditional measures such as M1 and M2 may be justified if the main objective is comparability with previous studies using these measures or corresponding panel studies. Thirdly, our study focused on a U.S. sample, expanding on prior evidence from the United Kingdom (Sturgis and Smith 2010) . While we expect similar findings in other populations, we lack direct evidence to support this claim. The lack of interpersonal comparability within a "homogeneous" sample of U.S. citizens may be amplified when comparing individuals from different cultures, countries, and languages. Nevertheless, we must exercise caution in generalizing our conclusions to other samples. Fourthly, the main aim of this study was to examine established measures as they have been used for decades. This implied that we use original wordings characterized by answer scales of different lengths (e.g., 4pt and 7pt). Although we assume scale length does not significantly affect our main variable of interest (i.e., shares of associations), a potential full-factorial design (7×2) where all seven items are measured with both scales, could explore any subtle differences in greater detail. Also, we used a particular set of emerging measures (i.e., IST (Robbins 2019 (Robbins , 2021)) ), and considering other emerging measures, such as the Risk Aversion question in the GSOEP and the UK Household Longitudinal Study, 20 could provide valuable insights. Fifth, we employed a probing technique (see "Experimental Design" section) that restated the trustee category originally presented (e.g., "In answering the previous question, who came to your mind when you were thinking about 'most people'?"). Repeating this category could be regarded as a form of priming potentially creating demand effects. For future research, exploring various probing strategies and utilizing designs that provide respondents with as little information as possible, and thereby avoiding any priming, could be a valuable avenue to pursue.

Finally, an open question emerges concerning whether frames of reference are systematically linked to respondents' demographic characteristics. Preliminary correlational evidence (see Online Appendix A.7) seems to show that this is not the case. This is encouraging and could mean that associations are predominantly random. However, to gain further clarity, future studies could extend the set of covariates considered and potentially employ a randomized design that attempts to induce associations of a particular kind to avoid post-hoc rationalization. 17. It may be beneficial to explore the semantic meaning of the term "stranger" and consider situations where individuals might perceive acquaintances as strangers for specific trust decisions, such as lending money. This highlights the situative nature of trust, where perceptions may vary depending on the context of the interaction (cf. Hardin 2002:9) . 18. For more detailed considerations between shorter and longer versions of IST, we refer readers to Robbins (2022) . 19. This approach could extract an individual specific general personal component of trust while acknowledging trust to be inherently situational, mitigate the effects of nonvalid associations in single items and provide a more robust assessment of trust across diverse situations. A high-truster would then be someone who has a high-level of trust across a large set of situations that involve trust. 20. "Are you generally a person who is fully prepared to take risks in trusting strangers or do you try to avoid taking such risks?".

References

1. resulting in additional 2×1, 500 entries M Out of 10,500 answers to trust questions, 3,000 responses were not probed
2. Detecting sentiment proves more complex than spotting mentions of known and unknown others due to several factors such as ambiguous word meanings
3. Figure 3. Neural signatures for conscious perception and maintenance in working memory. Couper Schonlau 10.7554/elife.23871.006 general, automated categorization is shown to result in meaningful time savings as opposed to eLife Sciences Publications, Ltd 2016. 2016 manual classification as soon as the data to be classified exceeds 1,500 documents (Schonlau and Couper
4. Fig. 1. Comparison of three groups across gender (groups are comparable). Gender — variable shows gender: 1 — male, 2 — female. CPBType — grouping variable. Col Pct — percent across columns. Frequency missing — missing data. Prob — probability, 0.6729 > 0.05 when statistical insignificance. Effective Sample Size — number of observations that were used for statistical analysis 10.17816/maj34127-30404 null ECO-Vector LLC (Publications) these instances were all coded as "unknown others
5. Gosain and Sardana 2017 the minority class to address the problem of class imbalance. This however did not lead to any further significant improvements Still we attempted oversampling (see e.g. Results are available upon request
6. Figure 8. Temporal structure during a sleep period following object-location memory task. Moreover 10.7554/elife.34467.020 SD = 0.27) we investigated the full dataset via paired sample t-tests with a Bonferroni adjusted alpha level of .016 per test eLife Sciences Publications, Ltd null 33 t(1,464) = 13.81, p < 0.001. Furthermore, but to a lesser extent (as is also depicted in Figure 2), M1, on average, results in higher trust scores than M2 (M = 0.4, SD = 0.26), t(1,475) = 3.11, p < 0.01. Also, the differences in trust scores for M2 and M3 are statistically significant, t(1,455) = 15.15, p < 0.001
7. To address potential outliers in individual situations, we propose exploring the concept of "cross-situational trust" (Bauer and Freitag 2018) and computing an average across measures. This approach could help mitigate the impact of strong outliers from specific situations
8. Online Appendix A.5.2 shows these results using data from the manually coded share of data only (n = 1,000/1,500) 1 500 Online Appendix A.5.3 shows these results using data for Subset 2 only
9. Age (catgeorical), sex, ethnicity, socioeconomic status, income and education
10. we randomized respondents to trust measures in Blocks #1 and #2; hence, we can conclude that the differences in the distribution of associations are the result of divergent frames evoked by the questions in respondents' minds 2010 Analogous to Sturgis and Smith
11. ANES: Codebook Variable Documentation (Version 04), 2000 Pilot Study (2000.PN) Anes References 2000
12. Formr: A Study Framework Allowing for Automated Feedback Generation and Complex Longitudinal Experience-Sampling Studies Using R Ruben C Arslan P Matthias Cyril S Walther Tata Behavior Research Methods 52 1 2020
13. Trust and Expected Trustworthiness: Experimental Evidence From Zimbabwean Villages Abigail Barr The Economic Journal 113 489 2003
14. Measuring Trust Paul C Bauer Markus Freitag The oxford handbook of social and political trust E M Uslaner Oxford University Press 2018
15. Asking Probing Questions in Web Surveys: Which Factors Have an Impact on the Quality of Responses? Dorothée Behr L Kaczmirek W Bandilla Michael Braun Social Science Computer Review 30 4 2012
16. Web Probing-Implementing Probing Techniques From Cognitive Interviewing in Web Surveys With the Goal to Assess the Validity of Survey Questions Dorothée Behr Katharina Meitinger Michael Braun Lars Kaczmirek 2017 GESIS -Survey Guidelines Mannheim GESIS -Leibniz-Institute for the Social Sciences
17. Analyzing Quantitative Data Norman Blaikie 2003 SAGE Publications Ltd London
18. Individual-Level Evidence for the Causes and Consequences of Social Capital John Brehm Wendy Rahn American Journal of Political Science 41 3 999 1997
19. An Experiment on the Effects of Embeddedness in Trust Situations: Buying a Used Car Vincent Buskens Jeroen Weesie Rationality and Society 12 2 2000
20. Does Travel Broaden the Mind? Breadth of Foreign Experiences Increases Generalized Trust Cao Adam D Jiyin William W Galinsky Maddux Social Psychological and Personality Science 5 5 2014
21. Ten Common Misunderstandings, Misconceptions, Persistent Myths and Urban Legends About Likert Scales and Likert Response Formats and Their Antidotes James Carifio J. Rocco Perla Journal of Social Sciences 3 3 2007
22. Foundations of Social Theory James S Coleman 1994 Harvard University Press Boston, MA
23. in Trust and reciprocity: Interdisciplinary lessons for experimental research Karen S Cook Robin M Cooper E. Ostrom and J. Walker 2003 Russell Sage Foundation New York Experimental Studies of Cooperation, Trust, and Social Exchange
24. Karen S Cook Russell Hardin Margaret Levi Cooperation Without Trust? New York Russell Sage Foundation 2005
25. Predicting Cross-National Levels of Social Trust: Global Pattern or Nordic Exceptionalism? Jan Delhey Kenneth Newton European Sociological Review 21 4 2005
26. How General Is Trust in 'Most People'? Solving the Radius of Trust Problem Jan Delhey Kenneth Newton Christian Welzel American Sociological Review 76 5 2011
27. The Radius of Trust Problem Remains Resolved Jan Delhey Kenneth Newton Christian Welzel American Sociological Review 79 6 2014
28. BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding Jacob Devlin Ming-Wei Chang Kenton Lee Kristina Toutanova J. Burstein, C. Doran, and T. Solorio 2019 Association for Computational Linguistics 1 Minneapolis, Minnesota in Proceedings of the 2019 conference of the north American chapter of the association for computational linguistics: Human language technologies
29. Upbringing, Early Experiences of Discrimination and Social Identity: Explaining Generalised Trust among Immigrants in Denmark Peter Thisted Dinesen Thisted 10.1111/j.1467-9477.2009.00240.x Scandinavian Political Studies 0080-6757 1467-9477 33 1 2010 Wiley
30. Does Generalized (Dis)Trust Travel? Examining the Impact of Cultural Heritage and Destination-Country Environment on Trust of Immigrants Peter Dinesen Thisted Political Psychology 33 4 2012
31. The Civic Personality: Personality and Democratic Citizenship Peter Dinesen Asbjørn Thisted Robert Sonne Nørgaard Klemmensen Politische Studien 62 1 2013 _suppl
32. Ethnic Diversity and Social Trust: Evidence From the Micro-Context Peter Dinesen Kim Thisted Sønderskov Mannemar American Sociological Review 80 3 2015
33. Do Strong Family Ties Inhibit Trust? John Ermisch Diego Gambetta Journal of Economic Behavior & Organization 75 3 2010
34. Measuring People's Trust John Ermisch Diego Gambetta Heather Laurie Thomas Siedler S C Noah Uhrig Journal of the Royal Statistical Society 172 4 2009
35. Machines That Learn How to Code Open-Ended Survey Data Andrea Esuli Fabrizio Sebastiani International Journal of Market Research 52 6 2010
36. A Nation-Wide Laboratory. Examining Trust and Trustworthiness by Integrating Behavioral Experiments Into Representative Surveys Ernst Fehr Urs Fischbacher Jürgen Bernhard Von Rosenbladt Gert G Schupp Wagner Journal of Contextual Economics-Schmollers Jahrbuch 122 4 2002
37. Spheres of Trust: An Empirical Analysis of the Foundations of Particularised and Generalised Trust Markus Freitag Richard Traunmüller European Journal of Political Research 48 6 2009
38. Automating Survey Coding by Multiclass Text Categorization Techniques Daniela Giorgetti Fabrizio Sebastiani Journal of the American Society for Information Science and Technology 54 14 2003
39. Do Social Connections Create Trust? An Examination Using New Longitudinal Data Jennifer L Glanville M A Andersson P Paxton 10.1093/sf/sot079 Social Forces Social Forces 0037-7732 1534-7605 92 2 2013 Oxford University Press (OUP)
40. How Do We Learn to Trust? A Confirmatory Tetrad Analysis of the Sources of Generalized Trust Jennifer L Glanville Pamela Paxton Social Psychology Quarterly 70 3 2007
41. Consequences of Failure to Meet Assumptions Underlying the Fixed Effects Analyses of Variance and Covariance Genev Glass D Percy James R Peckham Sanders Review of Educational Research 42 3 1972
42. Handling Class Imbalance Problem Using Oversampling Techniques: A Review Anjana Gosain Saanchi Sardana computing, communications and informatics 2017. 2017 ICACCI
43. Automated Classification for Open-Ended Questions With BERT Hyukjun Gweon Matthias Schonlau Journal of Survey Statistics and Methodology 2023 smad015
44. Russell Hardin Trust and Trustworthiness New York Russell Sage Foundation 2002
45. The Problem of Forming Social Capital. Why Trust Francisco Herreros 2004 Springer New York
46. The Rise and Fall of Social Cohesion: The Construction and De-Construction of Social Trust in the US, the UK, Sweden and Denmark Christian Larsen Albrekt 2013 OUP Oxford; Oxford
47. Measuring Generalized Trust Sebastian Lundmark Mikael Gilljam Stefan Dahlberg 10.1093/poq/nfv042 Public Opinion Quarterly PUBOPQ 0033-362X 1537-5331 80 1 2016 Oxford University Press (OUP)
48. Cost-Sensitive BERT for Generalisable Sentence Classification with Imbalanced Data Harish Madabushi Elena Tayyar Michael Kochkina Castelle arXiv 2003.11563 2020
49. Visual Design and Cognition in List-Style Open-Ended Questions in Web Probing Katharina Meitinger Tanja Kunz Sociological Methods & Research 0 0 2022
50. Stranger-Danger: What Do Children Know? Ellen Moran David Warden Lindsey Macleod Gillian Mayes John Gillies Child Abuse Review: Journal of the British Association for the Study and Prevention of Child Abuse and Neglect 6 1 1997
51. Measuring Trust: Experiments and Surveys in Contrast and Combination Michael Naef Jürgen Schupp 2009
52. What Have We Learned About Generalized Trust, If Anything? Peter Nannestad Annual Review of Political Science 11 1 2008
53. Open-Ended Versus Closed Probes: Assessing Different Formats of Web Probing Cornelia E Neuert Katharina Meitinger Dorothée Behr Sociological Methods & Research 52 4 2021
54. Prolific.ac-A Subject Pool for Online Experiments Stefan Palan Christian Schitter Journal of Behavioral and Experimental Finance 17 2018
55. Diversity, Social Capital, and Cohesion Alejandro Portes Erik Vickstrom Annual Review of Sociology 37 1 2011
56. Making Democracy Work: Civic Traditions in Modern Italy Robert D Putnam Robert Leonardi Raffaella Y Nanetti 1994 Princeton University Press Princeton
57. Measuring Generalized Trust: Two New Approaches Blaine G Robbins Sociological Methods & Research 51 1 2019
58. An Empirical Comparison of Four Generalized Trust Scales: Test-Retest Reliability, Measurement Invariance, Predictive Validity, and Replicability Blaine G Robbins Sociological Methods & Research 0 0 2021
59. Valid and Reliable Measures of Generalized Trust: Evidence From a Nationally Representative Survey and Behavioral Experiment Blaine G Robbins Socius 9 2022
60. Misanthropy and Political Ideology Morris Rosenberg American Sociological Review 21 6 1956
61. Trust in Social Relations Oliver Schilke Martin Reimann Karen S Cook Annual Review of Sociology 47 1 2021
62. Semi-Automated Categorization of Open-Ended Questions Matthias Schonlau Mick P Couper Survey Research Methods 10 2 2016
63. Race and Trust Sandra Susan Smith Susan 10.1146/annurev.soc.012809.102526 Annual Review of Sociology Annu. Rev. Sociol. 0360-0572 1545-2115 36 1 2010 Annual Reviews
64. Does Generalized Social Trust Lead to Associational Membership? Unravelling a Bowl of Well-Tossed Spaghetti Kim Sønderskov Mannemar European Sociological Review 27 4 2011
65. Trusting Strangers-The Concept of Generalized Trust in Perspective Dietlind Stolle Österreichische Zeitschrift für Politikwissenschaft 31 4 2015
66. Regression-Based Response Probing for Assessing the Validity of Survey Questions Patrick Sturgis Ian Brunton-Smith Jonathan Jackson Advances in questionnaire design, development, evaluation and testing John Wiley & Sons, Inc 2019
67. Assessing the Validity of Generalized Trust Questions: What Kind of Trust Are We Measuring? Patrick Sturgis Patten Smith International Journal of Public Opinion 22 1 2010
68. Piotr Sztompka Trust: A Sociological Theory Cambridge Cambridge University Press 1999
69. Identifying Social Trust in Cross-Country Analysis: Do We Really Measure the Same? Lars Torpe Henrik Lolle Social Indicators Research 103 3 2011
70. Eric M Uslaner The Moral Foundations of Trust Cambridge Cambridge University Press 2002
71. Measuring Social Capital: Orthodoxies and Continuing Controversies Van Deth International Journal of Social Research Methodology 6 1 Jan W. 2003
72. Experimental Evidence on Trust and Punishment Björn Vollan The Difference Between Kinship and Friendship: (Field-) 2011 40
73. Trust and Commitment in the United States and Japan Toshio Yamagishi Midori Yamagishi Motivation and Emotion 18 2 1994
74. Cross-Cultural Differences in Relationship-and Group-Based Trust Yuki William W Masaki Marilynn B Maddux Kosuke Brewer Takemura Personality & Social Psychology Bulletin 31 1 2005
75. she explores generalized social trust, specifically investigating the analysis of openended survey responses using automated approaches such as machine learning techniques
76. Deep Spatio-Temporal Learning for Parcel-Level Tea Quality and Yield Prediction Using IoT and Remote Sensing Data Lena Paul Mia Katharina 10.71448/bcds2562-2 Bulletin of Computer and Data Sciences Bull. Comput. Data Sci. 3072-2926 6 2 TechForum Publishing Group University of Freiburg and the Ludwig Maximilian University of Munich