Deliberation contrasts with other forms of communication because it involves reasoning about the arguments in the discussion (Bächtiger et al. 2018; Fishkin & Mansbridge 2017). Participants listen to what others think, consider the pros and cons, and decide what be the best for the community. Participants do not act out of prejudices but engage in rational reasoning during deliberation. Classically, deliberative ideals rest on reason, which reason would compete for the common good. Contemporary deliberation scholars have moved beyond and incorporated the role of self-interest and mutual justifiability. Nevertheless, the field still underscores the importance of providing reasoned justifications (Mansbridge, 2015).

Scholars have noted that this persistent emphasis on rational justifiability can inadvertently privilege those already adept at formal discourse, thus raising concerns about whose voices truly matter in deliberation. Sanders (1997) put it, ‘Some citizens are better than others at articulating their arguments in rational, reasonable terms. […] taking deliberation as a signal of democratic practice paradoxically works undemocratically’ (pp. 348–349). This is because deliberation ‘was governed by protocols of style and decorum that were themselves correlates and markers of status inequality’ (Fraser 1990, p.63). Those protocols have prevented citizens of a particular race, gender, and class from being equal. Pointing out that norms of deliberation ‘privilege speech that is formal, and general,’ Young (1997) argued that this type of speech is not innately universal. Lupia and Norton (2017) also pointed out that when people speak, they carry their social status, creating differences in the significance of the arguments based on the level of privilege. Therefore, social hierarchies project inequality into deliberation.

Whether social status affects an individual’s ability to articulate themselves in deliberation, deliberative ability, has been investigated in three ways: examining deliberation’s multifaceted ideals, measuring the space an individual occupies during discussion, and assessing an individual’s impact on group decision-making. So far, empirical studies seem to support critics’ concerns that deliberation advantages those who already hold social advantage and thus is undemocratic.

First, holistic measures that consider deliberative ideals correlate positively with social advantage. The Deliberative Quality Index (DQI) evaluates speech based on participation equality, level of justification, orientation toward the common good, respect among participants, constructive politics (reaching consensus), and reciprocal respect (Bächtiger, Gerber & Mm Fourner-Tombs, 2022; Steiner 2012). In Europolis, a transnational Deliberative Poll on European Union policy proposals, working-class participants scored lower on the DQI mainly for not mentioning the global economy—an indicator of the common good—rather than lacking justification (Gerber et al. 2018). Comparisons between citizens and political elites similarly imply that one’s social position increases familiarity, since politicians scored higher on the DQI than citizens (Jennstål 2019; Mockler 2022; Pedrini 2014).

Second, the less advantaged speakers appear less important. Males are represented more often than females. For instance, in an experimental, unmoderated political deliberation, females spoke more than males only when they had a critical mass, or when the groups voted unanimously (Karpowitz, Mendelberg & Shaker 2012). Moreover, participants with greater resources tend to participate more actively (Himmelroos, Rapeli & Grönlund 2017). Observations from local townhall meeting in India also support women faced disadvantages in agenda setting (Parthasarathy, Rao & Palaniswamy 2019). Furthermore, identical arguments from a female-passing character carried less authority than those from a male-passing character, illustrating how females remain underrepresented (Beauvais 2021).

Third, critics fear that the socially less advantaged participants may be dominated by the more advantaged when opinions form (Mansbridge 2010; Sanders 1997; Young 1990). Rather than weighing the merits of the arguments themselves, less advantaged participants might rely on the speaking skills of their privileged counterparts, distorting public will formation. Lupia and Norton (2017) explained: ‘The communicative acts that precede the outcome will use language that conveys power. They will be used by people who are more and less skilled in using language to acquire power. If participants are not paying close attention to these skill imbalances, and if the deliberative rules are not built to mitigate deleterious effects of such imbalances, participants are likely to be swayed by the skilled’ (p. 73). Indeed, well-educated group leaders strongly influenced less educated participants in an unstructured public deliberation in sub-Saharan Africa (Humphreys, Masters & Sandbu 2006). Similarly, jurors of color changed their opinions more frequently than White jurors, altering the outcome of the jury verdicts (Karpowitz et al. 2024). These findings suggest that participants with higher status often dominate opinion formation.

Deliberation has not been equal across participants. Historically empowered groups—men, Whites, and higher class—tend to speak more and receive greater representation in deliberation than historically marginalized groups. Thus, deliberation is not as democratic as it aspires to be. This study examines whether social status translates into deliberative advantage regarding how individuals deliberate and how opinions form.

Data for the ‘Shaping Our Future’ Deliberative Poll,1 held on May 1–2, 2021, were provided by Stanford Deliberative Democracy Lab. The event included 664 young American adults to address growing importance of young voters in America. Participants were primarily recruited from 35 postsecondary institutions, with additional members from Generation Lab to diversify educational backgrounds; hence, most of the sample was a convenience sample, and not representative of the national young voters (Stanford Deliberative Democracy Lab 2021). Of these participants, 623 provided demographic data, and 611 completed opinion and knowledge surveys. The number of participants completing all procedures was 589 on Day 1 and 544 on Day 2. The event took place on an online platform developed by Stanford Crowdsourced Democracy Team and Stanford Deliberative Democracy Lab, featuring automated moderation and transcription.

The Deliberative Poll followed a standard seven-step process: recruitment, pre-deliberation survey, distribution of briefing materials, random assignment to small groups (based on arrival time), small-group discussions, plenary sessions with expert responses, and post-deliberation survey. On Day 1, participants deliberated on three electoral reform topics, and on Day 2, discussions focused on the Civilian Climate Corps and economic inequality. All topics were covered in both small-group discussions and plenary sessions. Participants chose whether to attend both days without prior knowledge of the specific agendas. Outlines of agendas are in Appendix A.

Table 1 summarizes the demographic characteristics of the 623 participants who provided their information. The sample comprised 56% women (n = 349), 39% men (n = 246), and 4.5% non-binary or other genders. In terms of education, 74% (n = 464) were more educated, with 47% (n = 295) pursuing a bachelor’s degree and 27% (n = 169) holding or pursuing higher degrees. The Spearman correlation analysis revealed that younger participants did not necessarily lack educational attainment (ρ = –0.02, p = 0.55). The racial composition included 45% White participants (n = 282), 20% Asian (n = 122), 13% Black (n = 82), 11% Latinx (n = 68), and 11% from other racial backgrounds (n = 69). Participants were distributed across 60 groups on Day 1 and 56 groups on Day 2, each averaging 10 participants, reflecting the overall demographic distribution. The group characteristics are in Table 2.

Deliberative advantage. Drawing on the literature, this research measured deliberative advantage using four variables: speaking time, speaking turns, number of justified speech acts, and sophisticated language use (e.g., Dutwin 2003; Karpowitz, Mendelberg & Shaker, 2012; Siu 2017). Each variable is detailed below.

Speaking time. The online platform tracks the total time a participant spent speaking each day after pressing the ‘Request to Speak’ or ‘Interrupt’ button, provided the speech was recorded through the microphone and was audible to others. Speaking time is measured in seconds. Participants with room assignment logs but no recorded speech acts were coded as 0.

Speech turns. The total number of speech turns per day logged if the speech was audible to others by the online platform was used to measure this variable. Participants with log records of room assignments but no recorded speech were coded as 0.

Number of justified speech acts. Adapting from DQI’s Justification category, I counted the number of speech acts per speaker that contained relevant arguments extracted by DeliBERT, a machine-learning program designed to categorize arguments related to the proposals (Siu, Joseph & Hu 2024). Co-created by Stanford Deliberative Democracy Lab and Fileread Inc., a startup specializing in legal documents, DeliBERT was initially trained on presidential debates using Bidirectional Encoder Representations from Transformers (BERT) to extract premises and claims. It then employs Generative Pre-trained Transformer-4 (GPT-4) to summarize the extracted premises and claims according to researcher-provided proposals, categorizing each argument as supporting or opposing the proposal. Siu, Joseph and Hu (2024) reported that using transcripts from an Electoral Reform Deliberative Poll (Voices of Future), DeliBERT could annotate speech acts (F1 = 0.76) as effectively as human coders (F1 = 0.79), but faster.

In its most recent version, which recognizes story-telling as a valid form of justification, the DQI evaluates the quality of reasoning on a six-point scale, from 1 (no reasoning) to 6 (two or more logically linked reasons) (Steiner 2012). However, dismissing a strong single argument in favor of multiple weaker ones can be misleading (Chen 2021), since one argument may outweigh several lesser arguments if it addresses a critical value. Accordingly, I used DeliBERT’s determination of whether an argument supports or opposes a proposal to measure the linkage between justification and claim. Participants assigned to a small group but who never spoke were coded as 0. Examples of DeliBERT’s assessments, including how it judges argument relevance to the topic, appear in Appendix B.

Sophisticated language use. The degree of sophisticated language use was determined with the analytical thinking score (hereafter analytical thinking) from the Linguistic Inquiry and Word Count (LIWC) 22 program, a dictionary developed to assess psychometrics of a given text (Pennebaker et al. 2022). Analytical thinking is defined as ‘a deliberate mode of thought wherein complex concepts are deconstructed into more manageable components and their interrelations’ (Jordan et al. 2019, p.3477). When people produce analytical writings, they tend to use more articles and prepositions to incorporate different concepts. In contrast, communicating in an informal, personable style, people tend to use more pronouns, adverbs, negations, auxiliary verbs, and conjunctions (Jordan et al. 2019; Pennebaker et al. 2014). Reflecting psychological research, LIWC 22 summarizes analytical thinking as a single variable based on the relative use of function words in each corpus (Jordan et al. 2019; Pennebaker et al. 2022). Therefore, analytical thinking measures sentence complexity in terms of logical and formal structure, addressing the concerns raised by critics of deliberation.

Analytical thinking is comparable to cognitive complexity (CC), a construct often used in deliberation research (e.g., Brundidge et al. 2014; Wyss, Beste & Bächtiger 2015). CC refers to the ability to consider multiple perspectives, discern ideas, and integrate them (Houck et al. 2022; Wyss, Beste & Bächtiger 2015). However, the measurement commonly used in deliberation literature, suggested by Wyss et al. (2015), relies on LIWC2007 categories that are now deprecated due to low internal reliability (Pennebaker et al. 2015). Consequently, this research uses analytical thinking introduced in LIWC in 2015 to measure cognitive abilities reliably (Markowitz 2023).

Scores from each speech act were aggregated to produce a mean analytical thinking score per person, capturing each individual’s average ability to speak in a sophisticated manner. Each score was divided by 100, so the scores close to 1 suggest formality, whereas values close to 0 imply personable language use. Participants who never spoke were coded as 0. Examples of speech acts with low and high analytical thinking appear in Appendix B.

Pressure. Pressure indicates the burden a participant feels after they learn their opinion is different from those of vocal participants. I measured pressure as the pre-deliberation opinion difference between each participant and the top 25% of group members, ranked by their expressiveness. Expressiveness is the average of four standardized indicators—speaking time, speech turns, justified arguments, and language sophistication.

Socially advantaged status. Building on literature (e.g., Beauvais 2021; Luskin et al. 2022; Sanders 1997; Siu 2017), socially advantaged status is defined as being male, White, and more educated. Since income level was loosely defined in the survey,2 education was used as a proxy for social class. To account for the high proportion of participants who were in college at the time of study, this research uses them as the baseline group for education. Participants enrolled in education beyond college are viewed as having greater advantages, while those with less education than the baseline will be viewed as having lower advantages.

Attitudes about proposals. All responses were collected via the online survey tool Qualtrics, where participants accessed surveys and self-administered their opinions on computers or mobile phones. They indicated their attitudes toward each proposal using an 11-point Likert scale, ranging from 0 (Strongly Oppose) to 10 (Strongly Favor), with 5 signifying a neutral position. Participants could also select ‘No Opinion’ option, and they were treated as missing data.

Domination. If a group’s post-deliberation average opinion about an issue went to the same direction of the group’s pre-deliberation average opinion of the advantaged participants, the group was regarded to be ‘dominated’ on that issue. The advantaged participants were identified with eleven combinations of socially advantaged status. The mathematical notation on the measurement is outlined in Appendix D.

Hypothesis 1, which posits that socially advantaged participants have the greater deliberative advantage, was examined using mixed-level zero-inflated Poisson (ZIP) models and a zero-inflated Beta (ZI-Beta) model with socially advantaged status due to the deliberative advantages’ distributional characteristics and the excess zeros in counts (ZIP) and proportions (ZI-Beta). Excess zeros refer to zeros that do not follow the expected distribution and can result from structural, research design, observer, and random errors3 (Zuur et al. 2009).

Zero-inflated models first separate zeros and non-zeros and estimate zeros via logistic regression. Then, they fit Poisson or Beta regression to the non-zero cases. In this study of speech time, excess zeros are cases that do not conform to a Poisson distribution, possibly due to structural barriers such as gender, racial, and educational disparities. The ZIP model is particularly useful for simultaneously modeling count data and the excess zeros that arise from underlying mechanisms (Atkins & Gallop 2007).

The mixed-level ZIP and ZI-Beta models account for each participant’s speaking behavior within their randomly assigned groups. Speech time, speech turn, number of justified speech acts, and the degree of sophisticated language use were each regressed on participant’s gender, race, and education levels. Mixed-level ZIP models estimated speech time, speech turn, and number of justified speech acts, and a mixed-level ZI-Beta model was applied to degree of sophisticated language use. Appendix C illustrates how zero-inflated models estimate these two components, with detailed equations for each model.

Hypothesis 2, which predicted that participants of higher social status would dominate their group’s opinions, was tested at the group level by calculating the percentage of group-issue combinations dominated by socially advantaged participants (Luskin et al. 2022).

Hypothesis 3, which examined whether participants feel pressured to conform to the opinions of articulate participants was investigated drawing on social conformity research (Sanders 2012), addressing the question, ‘Is the degree of opinion change dependent on the gap between one’s opinion and that of vocal participants?’ For participants who provided opinions (excluding ‘Don’t know’), their post-deliberation opinion change was regressed on both their pre-deliberation opinions and the social comparison pressure. Additionally, I examined how a participant’s advantage moderates this effect, testing the hypothesis that less advantaged participants are more prone to defer to the opinions of those with greater advantages. Appendix E contains the regression models used in this analysis.

All models were fitted using R version 4.4.1 (R Core Team 2024). Generalized linear mixed models (GLMMs) were estimated via the glmmTMB package (Brooks et al. 2017), and all results were visualized using ggplot2 (Wickham 2016). To compute Nakagawa pseudo R² values and the intraclass correlation coefficient (ICC),4 I employed the sjPlot package (Lüdecke 2023).

Mixed ZIP regression models and a ZI-Beta regression model were fitted to examine Hypothesis 1, which posits that socially more advantaged participants have greater deliberative advantage. Overall, the analyses confirmed that socially more advantaged participants exhibited more speaking time, speech turns, number of justified speech acts, and degree of sophisticated language use. Each model is described below.

To examine Hypothesis 2, which posits that more advantaged participants would shape group opinions, I compared post-deliberation attitudes and pre-deliberation attitudes for all group-issue combinations (n = 460). A group-issue combination was deemed ‘dominated’ if its post-deliberation opinion shifted toward the pre-deliberation attitudes of the more advantaged participants. Table 9 presents the count of group-issue combinations dominated by participants of specific establishment types, where relevant. The proportion of group-issue combinations dominated by the more advantaged participants ranged from 48.04% to 57.84%. Notably, among the groups that included participants who were male, White, and educated beyond an associate’s degree, 53.95% of the group-issue combinations were dominated.

The effect of exposure to, and pressure from, participants with deliberative advantages—characterized by high speaking time, speech turns, number of justified speech acts, and degree of sophisticated language use—on post-deliberation opinions was examined across eight agenda items over two days using linear regression models. The analysis considered both the direct effect of pressure and the moderating roles of participants’ social advantages, including gender, race, and education level. As shown in Table 10, the impact of pressure on post-deliberation opinions varied by agenda item.

Overall, the direct effect of pressure from expressive participants on post-deliberation opinions produced mixed results. Of the eight agenda items, two showed significant effects on opinion change. For example, during the discussion on ranked-choice voting (Agenda 3 of Day 1), pressure led participants to align their opinions more closely with those of high-deliberative-advantage participants (β^ = 0.22, t(479) = 2.27, p = 0.02) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[(\hat \beta \, = 0.22, t{(479) = 2}{.27,}\,p\, = 0.02)\] \end{document} . Conversely, in the discussion about prioritizing vulnerable communities for the Civilian Climate Corps (Agenda 2 of Day 2), pressure decreased support for the proposal (β^ = −0.34, t(430) = −2.41, p = 0.02) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ (\hat \beta \, = -0.34,t(430) = -2.41,\,p\, = 0.02) \] \end{document} . Across all agenda items, the most consistent direct predictor of post-deliberation opinion was each participant’s pre-deliberation opinion, which exhibited positive coefficients throughout.

To investigate whether the impact of pressure differed based on participants’ social advantages, moderating effects were analyzed. While demographic factors did moderate this relationship, patterns were not uniform across all agendas. Generally, the pressure–gender interaction was nonsignificant. However, being White positively moderated the link between pressure and post-deliberation opinion during Agenda 2 on Day 2 (β^ = 0.37, t(430) = 2.49, p = 0.01) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ (\hat \beta \, = 0.37,t(430) = 2.49,\,p\, = 0.01) \] \end{document} . This moderation effect is illustrated in Figure 5.

Lower educational attainment also produced varied responses to pressure. For instance, during discussions on funding the Civilian Climate Corps (Agenda 1 of Day 2), participants with lower education expressed less support for the proposal the more their opinions diverged from those of advantaged participants (β^ = −0.53, t(332) = −2.62, p = 0.009) \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ (\hat \beta \, = - 0.53,t(332) = - 2.62,\,p\, = 0.009) \] \end{document} , illustrated in Figure 6. In contrast, on regional minimum income (Agenda 3 of Day 2) and wealth tax (Agenda 5 of Day 2), these same participants felt stronger pressure to conform (β^ = 0.40, t(332) = 2.06, p = 0.04 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ (\hat \beta \, = 0.40,t(332) = 2.06,\,p\, = 0.04 \] \end{document} for Day 2 Agenda 3, β^ = 0.47, t(477) = 3.73, p = 0.000 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ \hat \beta \, = 0.47,t(477) = 3.73,\,p\, = 0.000 \] \end{document} for Day 2 Agenda 5), as shown in Figures 7 and 8 respectively.

The models that showed either a direct effect of pressure or a significant interaction between pressure and social advantage generally exhibited moderate explanatory power, with adjusted R² values ranging from 0.15 to 0.45 across the agenda items.