Theory, empirical evidence, and hypotheses

We define political trust as ‘a global affective orientation toward government, one that reflects citizens’ feelings about government performance, processes, and probity, as well as their partisan biases in a polarized era’ (Hetherington and Rudolph Reference Hetherington and Rudolph2022: 356). This definition underscores that trust is not purely cognitive but inherently involves affective orientations – a premise central to our study. It also highlights the role of partisanship, which occupies a distinctive position in trust formation: neither a strictly rational heuristic nor a purely affective response, partisanship functions as a hybrid factor that filters how citizens evaluate government. In the polarized U.S. context of our study, we account for this by including party preference as a covariate in our analyses.

‘Cognition-based’ and ‘affect-based’ trust

An influential tradition in trust research holds that trust originates from informed and rational judgments. This ‘cognitive-based’ approach to trust, in essence, suggests that individuals base their trust judgments on purposeful and thoughtful evaluations of objects that primarily take a cognitive form (Hooghe, Marien, and de Vroome Reference Hooghe, Marien and de Vroome2012; Metzger and Flanagin Reference Metzger and Flanagin2013). For example, for political trust, trustors would increase or decrease their trust based on knowledge and information they have about a political entity, from which they then derive predictions about the trustee’s future trustworthiness.Footnote ² Lahno (Reference Lahno2020) traces this cognitive definition of trust back to the 1980s and describes how this notion has found its way into contemporary trust research by the predominance of using dilemma games to measure (i.e., interpersonal) trusting behavior. Trust games typify situations of individual rational decision-making (Lahno Reference Lahno2020), and the key explanation of trusting behavior in these game-theoretical accounts is often achieved by assumptions of rational choice theory, where ‘the human agent will choose his actions rationally in the light of […] aims’ (Lahno Reference Lahno2020: 148). In sum, a cognitive-based understanding of trust aligns with the notion that trust judgments are made upon the basis of risk calculations and rational choice-making processes (Coleman Reference Coleman1994; Hardin Reference Hardin2002; Levi and Stoker Reference Levi and Stoker2000).

However, in addition to this cognitive-based form of trust, others have suggested that there is another, a so-called ‘affect-based’ type of trust, and various authors have long emphasized that trust includes both emotional and cognitive dimensions (e.g., Dunn and Schweitzer Reference Dunn and Schweitzer2005; Finucane, Alhakami, Slovic et al. Reference Finucane, Alhakami, Slovic and Johnson2000; Lahno Reference Lahno2020; Lee, Dirks, and Campagna Reference Lee, Dirks and Campagna2023; Lewis and Weigert Reference Lewis and Weigert1985; McAllister Reference McAllister1995; Midden and Huijts Reference Midden and Huijts2009; Myers and Tingley Reference Myers and Tingley2016; Robbins Reference Robbins2016; Theiss-Morse and Barton Reference Theiss-Morse and Barton2017). Grimmelikhuijsen, for instance, describes that ‘a decision to trust a government organization may […] not always be conscious and/or rational’ (Grimmelikhuijsen Reference Grimmelikhuijsen2012: 57). In affect-based trust, the basis for trust (or distrust) lies in emotional ties (McAllister Reference McAllister1995) and is further grounded in an individual’s attributions concerning the motives for the trustee’s behavior (McAllister Reference McAllister1995: 29). This conceptualization of trust seems to specify emotion and affect as a core part of the construct itself (Lee, Dirks, and Campagna Reference Lee, Dirks and Campagna2023), consistent with the definition of political trust introduced above (Hetherington and Rudolph Reference Hetherington and Rudolph2022). Bertsou (Reference Bertsou2019b, Reference Bertsou2019a) further argues that political distrust should be understood as a distinct concept rather than merely the absence of trust, involving its own affective and evaluative processes. This distinction is relevant because the emotions elicited during trust evaluations may differ qualitatively depending on whether respondents are expressing trust or distrust. It is also important to note that trust relationships can be horizontal (between individuals) or vertical (between citizens and institutions). Our study focuses on a single vertical trust judgment – trust in the federal government – which means that the affective content we detect reflects citizens’ orientations toward a specific institutional object rather than generalized interpersonal trust. Importantly, we do not assume that cognitive and affective routes to trust are mutually exclusive; rather, they likely co-occur, and open-ended survey answers may contain both cognitive reasoning and affective content simultaneously.

Before reviewing the empirical evidence, it is useful to clarify key terms. Affect is an umbrella term encompassing both moods and emotions (Lee, Dirks, and Campagna Reference Lee, Dirks and Campagna2023: 549). A mood (or sentiment) captures only valence – whether an affective state is positive, negative, or neutral – and tends to be diffuse and longer-lasting. Emotions, by contrast, are shorter in duration, more intense, and directed at specific objects or events (Dunn and Schweitzer Reference Dunn and Schweitzer2005: 737). Importantly, emotions arise from cognitive evaluations of situations, which is why different emotions can have distinct effects on attitudes and behavior (Smith and Ellsworth Reference Smith and Ellsworth1985).

Two major theoretical frameworks predict how specific emotions influence political attitudes, including trust. First, Affective Intelligence Theory (AIT) (Marcus, Neuman, and MacKuen Reference Marcus, Russell Neuman and MacKuen2000; Marcus, Valentino, Vasilopoulos et al. Reference Marcus, Valentino, Vasilopoulos and Foucault2019) distinguishes between the disposition system, which generates enthusiasm when habitual routines succeed, and the surveillance system, which triggers anxiety in response to novel threats and prompts deliberative information-seeking. Anger, by contrast, arises when identifiable actors are held responsible for threats, reinforcing reliance on existing dispositions rather than deliberation (Vasilopoulos, Marcus, Valentino et al. Reference Vasilopoulos, Marcus, Valentino and Foucault2019). AIT thus predicts that anxiety and anger have qualitatively different effects on political judgment. For trust specifically, enthusiasm should reinforce positive orientations toward government, anxiety may prompt reassessment of prior trust levels, and anger should strengthen reliance on existing dispositions – potentially deepening distrust among those already skeptical of government.

Second, Cognitive Appraisal Theory (e.g., Smith and Ellsworth Reference Smith and Ellsworth1985) holds that specific emotions emerge from how individuals evaluate situations along dimensions such as certainty, control, and responsibility attribution. In the trust literature, this has generated competing predictions: control appraisals suggest that anger (an other-person-control emotion) most strongly diminishes trust (Dunn and Schweitzer Reference Dunn and Schweitzer2005), while certainty appraisals predict that low-certainty emotions such as anxiety have the largest negative effect (Myers and Tingley Reference Myers and Tingley2016). Our study is, furthermore, closely inspired by Lodge and Taber’s (Reference Lodge and Taber2013) concept of ‘hot cognition.’ Their central claim is that political concepts (e.g., a party, a politician, a policy) are instantly and without intentional control tagged with an affective valence – classified as good or bad – based on prior associations stored in long-term memory (Lodge and Taber Reference Lodge and Taber2013: 44). This allows the brain ‘to use affect as real-time information to promote quick, efficient, spontaneous responses’ (Lodge and Taber Reference Lodge and Taber2013: 48). Applied to trust, hot cognition implies that when respondents evaluate how much they trust government, affective associations are activated automatically alongside deliberate reasoning, and these associations may shape the resulting judgment.

Previous empirical evidence and hypotheses

Various studies have examined the influence of emotions on trust (for an overview see Lee, Dirks, and Campagna Reference Lee, Dirks and Campagna2023), yet this literature has predominantly focused on interpersonal rather than political trust (cf. Grimmelikhuijsen Reference Grimmelikhuijsen2012; Theiss-Morse and Barton Reference Theiss-Morse and Barton2017). Theiss-Morse and Barton (Reference Theiss-Morse and Barton2017), for example, lament that the vast majority of previous research on political trust takes a cognitive approach, stressing that ‘ignoring emotions is detrimental to our efforts to get a full understanding of political trust’ (Theiss-Morse and Barton Reference Theiss-Morse and Barton2017: 160). However, recent studies have begun to address this gap. Drawing on AIT and Cognitive Appraisal Theory, Erhardt, Freitag, Filsinger et al. (Reference Erhardt, Freitag, Filsinger and Wamsler2021) show that fear and anger had divergent effects on trust in government during the COVID-19 pandemic, with fear increasing trust (a rally-round-the-flag effect) and anger decreasing it. Similarly, Meer, Steenvoorden, and Ouattara (Reference Meer, Steenvoorden and Ouattara2023) find that fear of infection was a key micro-level mechanism behind the COVID-19 rally effect on political trust. Vasilopoulos, Mcavay, Brouard et al. (Reference Vasilopoulos, Mcavay, Brouard and Foucault2023) demonstrate that emotions shaped both governmental trust and support for restricting civil liberties during the pandemic, while Albertson and Gadarian (Reference Albertson and Gadarian2015) provide a broader account of how anxiety influences democratic citizenship and political attitudes.

Building on these findings, we now review the interpersonal trust literature in more detail, as it provides the theoretical basis for our hypotheses. Dunn and Schweitzer (Reference Dunn and Schweitzer2005) explore the influence of emotions on interpersonal trust using survey experiments. They find that moods with positive valence increase trust, while moods with negative valence decrease it, consistent with the ‘affect-as-information’ model (Schwarz and Clore Reference Schwarz, Clore, Fiedler and Forgas1991) in which people attribute their current mood to the judgment they are evaluating (Dunn and Schweitzer Reference Dunn and Schweitzer2005: 737). Beyond valence, they find that happy participants were more trusting than sad participants, and sad participants were more trusting than angry participants. Through the lens of Cognitive Appraisal Theory, they explain these differences by the ‘cognitive nature’ of the respective emotion: anger, driven by ‘other-person control’ – attributing responsibility to others – has a larger negative effect on trust than sadness, which involves less other-person control (Dunn and Schweitzer Reference Dunn and Schweitzer2005: 738). Happiness, as an emotion of weak control appraisals, has a positive effect on trust.Footnote ³ Other empirical accounts argue for ‘certainty appraisals’ instead of ‘control appraisals’ as the key dimension (e.g., Albertson and Gadarian Reference Albertson and Gadarian2015; Myers and Tingley Reference Myers and Tingley2016). Myers and Tingley (Reference Myers and Tingley2016) find that negative emotions decrease trust only when they make people feel less certain about their current situation (e.g., anxiety). In contrast to Dunn and Schweitzer, they find no significant effect of anger – a high-certainty emotion – on trusting behavior. They argue that irrespective of other-person control, emotions with high certainty should have little impact on trust compared to emotions of low certainty such as anxiety. The authors note that differences in measurement approaches (survey items vs. trust games) may partly explain the divergent findings across studies.

In sum, both AIT and Cognitive Appraisal Theory generate predictions about distinct effects of specific emotions on trust, though they differ in which appraisal dimensions they emphasize (see above). What unites these accounts is the distinction between an emotion’s valence (i.e., sentiment) and the discrete emotion itself, each of which may independently shape trust judgments. Drawing on these theoretical accounts, we analyze both sentiment (valence) and discrete emotions in open-ended trust responses, providing a multi-dimensional assessment of their affective content. Our first hypothesis follows directly from the affect-based trust literature reviewed above:

In our second hypothesis, we aim to explore the link between these affective rationales and actual trust ratings. In line with previous research that investigated the impact of emotions on trust using survey items, we adopt an approach that acknowledges the valence (i.e., positive or negative) of an emotion as well as its ‘cognitive nature’. In particular, following Dunn and Schweitzer (Reference Dunn and Schweitzer2005), we assume that both sadness and anger, due to their negative valence, have a negative effect on political trust. However, due to the idea that anger is an emotion characterized by other-person control, again following Dunn and Schweitzer (Reference Dunn and Schweitzer2005), we assume that the effect of anger on trust is larger than that of sadness.Footnote ⁴ Happiness, due to its positive valence, should have a positive effect on trust. This expectation is also consistent with AIT, where enthusiasm (the closest analogue to happiness in our four-category taxonomy) reinforces positive orientations. We note, however, that our emotion categories (happy, sad, angry, neutral) do not fully map onto AIT’s core emotions (enthusiasm, anxiety, anger), particularly lacking a measure of anxiety/fear. This limits a direct test of AIT’s predictions but does not preclude testing its disposition system and anger predictions.

Analytical strategy

We analyze the open-ended answers in terms of whether and to what extent they exhibit affective components, including sentiment as well as emotions. Our multimodal design matches each analytical task to the modality where it can be most reliably performed: we use transcribed text for sentiment analysis, because valence classification (positive/negative/neutral) can be achieved with high accuracy from text alone using established Natural Language Processing (NLP) models; we use the original audio recordings for emotion detection, because discrete emotion recognition benefits substantially from paralinguistic cues (intonation, pitch, volume, pauses) that are lost in transcription (Lu, Cao, Zhang et al. Reference Lu, Cao, Zhang, Chiu and Fan2019). As a cross-modality robustness check, we also present text-based emotion results using the NRC word-emotion lexicon in the Online Appendix (Section A4.2). Put differently, we not only use traditional text-based methods for Sentiment Recognition but also Speech Emotion Recognition (SER). Both tasks have the objective of taking as input a spoken or written sentence and, as output, a classification of the expressed sentiment or emotion, such as neutral, negative, positive, happy, sad, and more.

Figure 1 illustrates the various steps in our workflow and the different approaches chosen to detect sentiment and emotions.

Figure 1.

Methods for Sentiment and Emotion Analysis in the context of our Analytical Strategy.

Note: Our Appendix additionally shows results where emotions were classified with transcribed documents using the NRC word-emotion lexicon (Section A4.2). Section A3.2 shows findings for a sentiment classification based on the transcribed documents using different dictionary approaches (AFINN and VADER).

Below, we describe in detail how these different steps depicted in Figure 1 are achieved.

Sentiment analysis

We start by analyzing the sentiment of the survey responses, classifying them into positive, negative, or neutral tones. In this study, we are exploring two approaches for this task; the first one is pysentimiento, a fine-tuned BERT model, and the second approach is zero-shot prompting with GPT-3.5. While both approaches constitute a deep-learning architecture, they, for example, differ with regard to whether they are fine-tuned, that is, specifically trained for the sentiment task or not.

BERT (Bidirectional Encoder Representations from Transformers) (Devlin, Chang, Lee et al. Reference Devlin, Chang, Lee and Toutanova2018) is a language model with weights that contain contextual word representations (i.e., word embeddings), and these weights can be fine-tuned for diverse natural language processing tasks. Fine-tuning is the process of ‘refining’ a pre-trained model on a smaller task-specific dataset to learn a specific downstream task (e.g., sentiment classification). Nowadays, several pre-trained BERT models are available (Chiorrini, Diamantini, Mircol et al. Reference Chiorrini, Diamantini, Mircoli and Potena2021), and in this study, we are using a fine-tuned BERT model called pysentimiento (Pérez, Rajngewerc, Giudic et al. Reference Pérez, Rajngewerc, Giudici, Furman, Luque, Alonso Alemany and Martínez2023). Pysentimiento is built on top of Hugging Face’s Transformers library (Wolf, Debut, Sanh et al. Reference Wolf, Debut, Sanh, Chaumond, Delangue, Moi and Cistac2019) and represents a Python multilingual toolkit for social NLP tasks. Pysentimiento’s models for Sentiment and Emotion Analysis, developed by Pérez, Rajngewerc, Giudici et al. (Reference Pérez, Rajngewerc, Giudici, Furman, Luque, Alonso Alemany and Martínez2023), are deployed on the Hugging Face hub and can be easily accessed through the library. In particular, in our study, we are using the available sentiment model, which uses BERTweet (a RoBERTa model trained on English tweets) as its base model and was further trained with a dataset that contains 61,900 tweets annotated for polarity detection (SemEval 2017 Task 4 Subtask 1; see Pérez, Rajngewerc, Giudici et al. Reference Pérez, Rajngewerc, Giudici, Furman, Luque, Alonso Alemany and Martínez2023 for details).

Our second approach to detect sentiment in the survey responses does not depend on fine-tuning but instead leverages the power of language models via zero-shot prompting. In general, prompting is a modern classification approach that came into being through the emergence of large language models such as GPT-3. Prompting can be further dissected into zero-shot or few-shot prompting. Both approaches were shown to match or even exceed the performance of typical human coders for a variety of classification tasks (Burnham Reference Burnham2023; Rytting, Sorensen, Argyle et al. Reference Rytting, Sorensen, Argyle, Busby, Fulda, Gubler and Wingate2023). Zero-shot prompting is an approach that makes use of the ‘instructions contained in prompts without any training data’ (Chae and Davidson Reference Chae and Davidson2026: 3). Similarly, to sentiment dictionaries, zero-shot prompting ‘requires no training data and minimal programming to implement’ (Burnham Reference Burnham2023: 2); however, ‘[u]nlike sentiment dictionaries, it produces results comparable to, and sometimes better than, supervised classification’ (Burnham Reference Burnham2023: 2).

In our study, we are going to pursue zero-shot prompting with GPT-3.5-turbo, the successor to GPT-3, which, to date, is one of the largest existing language models (Brown, Mann, Ryder et al. Reference Brown, Mann, Ryder, Subbiah, Kaplan, Dhariwal, Neelakantan, Shyam, Sastry, Askell, Agarwal, Herbert-Voss, Krueger, Henighan, Child, Ramesh, Ziegler, Wu, Winter, Hesse, Chen, Sigler, Litwin, Gray, Chess, Clark, Berner, McCandlish, Radford, Sutskever and Amodei2020).Footnote ⁹ To achieve the classification of survey responses with GPT without providing human-annotated examples, we needed to provide the model with a specific prompt. Here, our goal was to provide a very clear and straightforward prompt that returns a single token (i.e., positive, negative, neutral) to detect the nature of associations in its most standard form: positive, negative, or neutral. Our prompt reads as follows (see Section A5 for more details on our prompt engineering):

where i is the survey response. We deliberately kept the prompt simple and did not provide GPT with an explicit definition of sentiment. Recent evidence suggests that large language models perform reliably on standard sentiment classification without elaborate prompting (Burnham Reference Burnham2023; Ziems, Held, Shaikh et al. Reference Ziems, Held, Shaikh, Chen, Zhang and Yang2023), and we validated our GPT classifications against a human-labeled subset, finding high agreement (see Section A3.1).

With both the BERT and the GPT approaches described above, we chose to pursue deep learning-based models due to the finding that other, more traditional and more simple approaches (i.e., dictionary approaches) exhibited very low accuracy in our sample. Historically, the analysis of sentiment in survey responses was achieved with theory-based dictionaries of affectively scored words (Mossholder, Settoon, Harris et al. Reference Mossholder, Settoon, Harris and Armenakis1995). Results, where we applied such dictionary methods to our data alongside a discussion of issues and challenges, can be found in Section A3.2.

Emotion analysis

The detection of emotions can be understood as a more difficult endeavor compared to sentiment analysis, ‘given the greater variety of classes and the more subtle differences between them’ (Chiorrini, Diamantini, Mircoli et al. Reference Chiorrini, Diamantini, Mircoli and Potena2021: 1). Automatic emotion recognition is a young research area, but we already know that recognizing emotions from text is challenging (for an overview of transformer models for emotion recognition from text, see Adoma, Henry, and Chen Reference Adoma, Henry and Chen2020; Kratzwald, Ilic, Kraus et al. Reference Kratzwald, Ilic, Kraus, Feuerriegel and Prendinger2018; and Nandwani and Verma Reference Nandwani and Verma2021, as well as Section A4.2 for an analysis with the EmoLex Dictionary) because it is stripped of all paralinguistic and acoustic features. We can assume that the inclusion of paralinguistic features such as intonation, pitch, volume, pauses, and also laughter or breathing noises (Lu, Cao, Zhang et al. Reference Lu, Cao, Zhang, Chiu and Fan2019) can be very helpful in recognizing emotional states in speech. This is why, in our study, for the task of emotion recognition, we are going to use methods from the field of Speech Emotion Recognition.

SER can be subordinated to the field of Automated Emotion Recognition and is a part of other disciplines such as Neurocomputing (Lope and Graña Reference Lope and Graña2023) and Affective Computing (Atmaja, Sasou, and Akagi Reference Atmaja, Sasou and Akagi2022; Madanian, Chen, Adeleye et al. Reference Madanian, Chen, Adeleye, Templeton, Poellabauer, Parry and Schneider2023). Speech Emotion Recognition can include traditional machine learning (for example, logistic regression) (Madanian, Chen, Adeleye et al. Reference Madanian, Chen, Adeleye, Templeton, Poellabauer, Parry and Schneider2023), but nowadays a large corpus of SER applications makes use of deep-learning architectures (Lope and Graña Reference Lope and Graña2023; Singh and Goel Reference Singh and Goel2022).

To achieve emotion detection in our survey answers, our study follows this trend of utilizing deep learning, and in particular, we utilize SpeechBrain (Ravanelli, Parcollet, Plantinga et al., Reference Ravanelli, Parcollet, Plantinga, Rouhe, Cornell, Lugosch and Subakan2021). SpeechBrain is an open-source deep learning toolkit built upon PyTorch specifically designed for speech and audio processing tasks. In general, SpeechBrain provides a comprehensive set of pre-built tools (‘recipes’) and models tailored for various speech-related tasks, including Speech Emotion Recognition.

For emotion detection, Wang, Boumadane, and Heba (Reference Wang, Boumadane and Heba2021) pre-trained a model that is based on the wav2vec 2.0 architecture. Wav2vec 2.0 (Baevski, Zhou, Mohamed et al. Reference Baevski, Zhou, Mohamed and Auli2020) was originally proposed for Automatic Speech Recognition but is nowadays also frequently used for speech emotion recognition (Chen and Rudnicky Reference Chen and Rudnicky2023; Pepino, Riera, and Ferrer Reference Pepino, Riera and Ferrer2021; Wang, Boumadane, and Heba Reference Wang, Boumadane and Heba2021). By training a wav2vec 2.0 model with data from the Interactive Emotional Dyadic Motion Capture (IEMOCAP) dataset, the authors achieve state-of-the-art results for different tasks.Footnote ¹⁰ For example, for the task of SER using the IEMOCAP public emotion recognition dataset, they achieved a weighted (averaged on 5 different seeds) accuracy rate of 79.58% across multiple emotions (model ‘PF-hbt-large’)Footnote ¹¹ . Similar performances for fine-tuned wav2vec 2.0 models for SER can be found in Chen and Rudnicky (Reference Chen and Rudnicky2023) (e.g., 74.3% unweighted accuracy for a wav2vec 2.0 model fine-tuned using a P-TAPT method). Wang, Boumadane, and Heba (Reference Wang, Boumadane and Heba2021) open-sourced their code within the SpeechBrain framework, which allows for easy implementation for researchers without extensive training resources.

SpeechBrain’s recipe for Speech Emotion Recognition classifies utterances into four emotions: anger, happiness, sadness, and neutrality (Wang, Boumadane, and Heba Reference Wang, Boumadane and Heba2021: 4). This decision of including four emotions is frequently found in SER research utilizing the IEMOCAP database and aims at ensuring consistency and comparability across other studies on the same database (Fayek, Lech, and Cavedon Reference Fayek, Lech and Cavedon2017). It is important to note that this four-category taxonomy does not fully align with the emotion categories central to political science theories. Most notably, it lacks fear and anxiety, two emotions that play a foundational role in Affective Intelligence Theory (e.g., Marcus Reference Marcus2003) and that have been shown to have distinct effects on political trust (Albertson and Gadarian Reference Albertson and Gadarian2015; Meer, Steenvoorden, and Ouattara Reference Meer, Steenvoorden and Ouattara2023). Under AIT, fear and anger generate opposing predictions for political attitudes, making the absence of a fear category a limitation that should be kept in mind when interpreting our emotion results. In general, research that includes data on linguistic and acoustic as well as paralinguistic features is still in its infancy. We are not aware of any studies that use the SpeechBrain emotion classifier, and social science research that takes advantage of such detailed language characteristics is not yet widespread. Noteworthy exceptions and applications lie within the field of political science (Dietrich, Hayes, and O’brien Reference Dietrich, Hayes and O’brien2019; Rittmann Reference Rittmann2024).Footnote ¹²

Our hypotheses will be investigated as follows. For our first hypothesis (H1) (‘Respondents, when requested to give reasons for their level of trust, include non-neutral sentiment (H1a) as well as emotional language (H1b) in their answers.’), we will examine the open-ended responses to determine which types of sentiment and emotion they include. As described above, sentiment will be measured with three categories (positive, negative, and neutral), and emotions will be measured with four categories (anger, happiness, sadness, and neutrality).

Our second hypothesis (H2), that the type of sentiment (H2a) or emotion (H2b) expressed in an open-ended survey answer impacts the closed-ended trust score, is analyzed by predicting the quantitative closed-ended measure of political trust with our sentiment and emotion variables. For this, we will use regression analysis and show findings both for bivariate models and multivariate models in which we control for age, education, socio-economic status, and party preference.