pith. sign in

arxiv: 2606.03786 · v1 · pith:ONQFZOZQnew · submitted 2026-06-02 · ⚛️ physics.soc-ph · stat.ME

Disentangling conviction and conformity: a Bayesian ideal point model of voting behaviour in online debates

Elena Candellone This is my paper

Pith reviewed 2026-06-28 07:47 UTC · model grok-4.3

classification ⚛️ physics.soc-ph stat.ME
keywords Bayesian ideal point modelonline debatesconvictionconformitypeer influencevoting behaviourDebate.orgmoral conviction
0
0 comments X

The pith

A Bayesian model of online debate votes finds conviction dominates on personal freedom topics while conformity dominates on moral issues like abortion and global warming.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a Bayesian logistic regression model inspired by ideal point models to separate conviction, driven by prior ideological beliefs, from conformity, driven by peer influence, in observed voting patterns. It applies the model to roughly 341,000 votes across 78,000 debates on 48 socio-political topics from Debate.org, first using large language models to infer topic and stance labels from debate text. The results show clear heterogeneity: conviction is stronger for lifestyle and personal freedom issues such as drug legalisation, while conformity is stronger for topics including abortion, gun rights, and global warming. A sympathetic reader would care because the split indicates when online opinions are anchored in personal priors versus shaped by the immediate social environment.

Core claim

The central claim is that voting behaviour in online debates exhibits substantial heterogeneity across topics in the relative strength of conviction versus conformity. Conviction, reflecting prior ideological beliefs, is the dominant driver on issues tied to personal freedoms and lifestyle choices such as drug legalisation and legalised prostitution. Conformity, reflecting peer influence, is the dominant driver on several topics widely regarded as cases of moral conviction, including abortion, gun rights, and global warming.

What carries the argument

Bayesian logistic regression model inspired by ideal point models, which estimates and quantifies the relative contribution of prior ideological beliefs versus peer influence to each observed vote.

If this is right

  • Votes on personal freedom topics such as drug legalisation reflect individual priors more than immediate group pressure.
  • Votes on moral topics such as abortion and global warming shift more readily with changes in the peer environment.
  • The stability of expressed opinions in online discourse varies systematically by topic according to which mechanism dominates.
  • Platform features that alter peer visibility would affect voting patterns differently across topic categories.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Platform designers could test interface changes that reduce peer cues on moral topics to increase the weight of conviction.
  • The same model applied to other debate sites might show whether the observed topic split is platform-specific or general.
  • Longitudinal tracking of individual users could test whether conformity effects produce lasting shifts in underlying beliefs.
  • Robustness checks using human-coded topic labels would directly test sensitivity to the LLM inference step.

Load-bearing premise

The Bayesian ideal point model structure together with the LLM-inferred topic and stance labels is sufficient to accurately separate and quantify the distinct contributions of prior beliefs versus peer influence.

What would settle it

Re-estimating the model after replacing the LLM topic labels with an alternative classification method or after altering the model specification produces a reversal or disappearance of the reported dominance patterns for the listed topics.

Figures

Figures reproduced from arXiv: 2606.03786 by Elena Candellone.

Figure 1
Figure 1. Figure 1: Platform structure and model description. Each debate (e.g. “XYZ should be legal”) has one PRO and one CON debater. Users can also vote PRO/CON. Each user has a belief vector of length K, with their self-reported stance on each of the K topics. Each debate has a topic vector of the same length, inferred from the debate text (see Section 2.2). I use these vectors to compute the topic alignment and peer infl… view at source ↗
Figure 2
Figure 2. Figure 2: Topic-specific parameter estimates. Posterior means and 90% credible intervals for βφ (peer influ￾ence) and βτ (topic alignment) across all 47 topics, grouped by issue. Bold labels indicate topics whose full 90% credible interval of the odds ratio exceeds e (i.e., β > 1) for the respective coefficient. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparing parameter estimates across issues. Each point plots the posterior mean of (βφ, βτ ) for a single topic; the surrounding ellipse represents posterior uncertainty. Points above the diagonal βτ = βφ are conviction-dominated; points below it are conformity-dominated; jointly-driven topics lie near the diagonal with both coordinates large. The dashed lines are the loci βφ = βτ , βφ = 1, and βτ = 1. 9 … view at source ↗
read the original abstract

Online debate platforms offer a unique window into the mechanisms driving opinion formation: they capture both explicit political preferences and the peer environment in which those preferences are expressed. In this work, I develop a Bayesian logistic regression model, inspired by ideal point models from political science, to disentangle two competing mechanisms of voting behaviour in online debates: conviction, driven by prior ideological beliefs, and conformity, driven by peer influence. I apply this framework to the Debate.org dataset, comprising approximately 341k votes across 78k debates on 48 socio-political topics. As the debate platform does not provide predefined topic labels for each debate, I infer the topic and stance from the debate text using large language models, and, with a Bayesian approach, I quantify the relative contribution of each mechanism. I find substantial heterogeneity across topics: conviction dominates on issues tied to personal freedoms and lifestyle choices, such as drug legalisation and legalised prostitution, while conformity dominates on several topics widely regarded as paradigmatic cases of moral conviction, including abortion, gun rights, and global warming. These results have implications for the stability of online political discourse and the design of deliberative platforms.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper develops a Bayesian logistic regression model inspired by ideal point models to disentangle conviction (prior ideological beliefs) from conformity (peer influence) in ~341k votes across 78k debates on Debate.org. Topics and stances are inferred via LLMs from debate text (no predefined labels exist). The model quantifies relative contributions per topic and reports substantial heterogeneity: conviction dominates on personal-freedom/lifestyle topics (e.g., drug legalisation), while conformity dominates on paradigmatic moral-conviction topics (abortion, gun rights, global warming).

Significance. If the separation of mechanisms holds, the work offers a scalable, quantitative approach to studying opinion formation in online platforms and challenges the assumption that topics like abortion or climate change are uniformly conviction-driven. The large dataset and Bayesian framing are strengths; reproducible code or parameter-free derivations are not mentioned.

major comments (3)
  1. [Methods] Methods (LLM topic/stance inference): No human validation, inter-annotator agreement, or sensitivity checks on the LLM labels are described. Because these labels produce the 48-topic partition that drives the headline conviction-vs-conformity heterogeneity, systematic label bias would propagate directly into the central claim.
  2. [Model and Results] Model identification and results: The conviction and conformity parameters are estimated from the same voting data used to report dominance; the manuscript should show (via equations or posterior checks) that the topic/stance covariates achieve genuine separation rather than post-hoc attribution of residuals. Without this, the reported topic-specific dominance lacks a clear falsification test.
  3. [Results] Results (topic heterogeneity): The classification of topics into conviction- or conformity-dominated groups should be accompanied by uncertainty measures (e.g., posterior probabilities or credible intervals on the relative contributions) so that borderline cases do not drive the narrative contrast with "paradigmatic moral conviction" topics.
minor comments (2)
  1. [Abstract] Abstract mentions the modeling approach but omits any equation or key identification assumption; a one-sentence statement of the logistic form would improve accessibility.
  2. [Methods] Notation for the two free parameters per user/debate should be introduced consistently when first defined.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major comment below and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Methods] Methods (LLM topic/stance inference): No human validation, inter-annotator agreement, or sensitivity checks on the LLM labels are described. Because these labels produce the 48-topic partition that drives the headline conviction-vs-conformity heterogeneity, systematic label bias would propagate directly into the central claim.

    Authors: We agree the manuscript lacks explicit validation of the LLM-inferred topics and stances. In revision we will add a new subsection reporting (i) a small-scale human annotation study on a random sample of debates to compute inter-annotator agreement and accuracy against LLM labels, and (ii) sensitivity checks across alternative prompts and models. These additions will quantify potential label bias and its propagation to the topic partition. revision: yes

  2. Referee: [Model and Results] Model identification and results: The conviction and conformity parameters are estimated from the same voting data used to report dominance; the manuscript should show (via equations or posterior checks) that the topic/stance covariates achieve genuine separation rather than post-hoc attribution of residuals. Without this, the reported topic-specific dominance lacks a clear falsification test.

    Authors: The model is a Bayesian logistic regression with user-level ideal-point parameters capturing conviction (prior ideological position) and debate-level peer-influence parameters capturing conformity; topic and stance enter as covariates that modulate the relative weight of each term. To demonstrate genuine separation we will insert the full generative equations and add posterior predictive checks plus simulation recovery experiments showing that the two mechanisms are identifiable and not merely partitioning residuals. These checks will serve as the requested falsification test. revision: yes

  3. Referee: [Results] Results (topic heterogeneity): The classification of topics into conviction- or conformity-dominated groups should be accompanied by uncertainty measures (e.g., posterior probabilities or credible intervals on the relative contributions) so that borderline cases do not drive the narrative contrast with "paradigmatic moral conviction" topics.

    Authors: We accept that uncertainty quantification is needed. We will revise the results section and figures to report, for each of the 48 topics, credible intervals on the relative contribution of conviction versus conformity together with the posterior probability that one mechanism dominates the other. This will allow readers to evaluate the robustness of the highlighted contrasts, including for abortion, gun rights, and global warming. revision: yes

Circularity Check

0 steps flagged

No significant circularity; model fit produces empirical estimates without self-referential reduction.

full rationale

The paper specifies a Bayesian logistic regression (ideal-point style) whose conviction parameter is a user-specific prior-ideology term and whose conformity term captures peer influence; both are estimated from the observed votes on LLM-labeled debates. The reported topic-level dominance findings are direct posterior summaries of those fitted parameters rather than quantities defined in terms of themselves or forced by construction. No self-citations, uniqueness theorems, or ansatzes imported from prior author work appear in the abstract or described derivation; the separation of mechanisms is imposed by model structure and the results are falsifiable against the same data. This is standard empirical modeling with no load-bearing circular step.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard Bayesian modeling assumptions plus the domain assumption that voting behavior can be decomposed into conviction and conformity components via the ideal point framework; LLM topic inference is treated as a measurement step without independent validation.

free parameters (2)
  • conviction and conformity parameters in logistic regression
    Parameters estimated from the 341k votes to quantify relative contributions of each mechanism.
  • ideal point locations for users and debates
    Latent positions fitted within the Bayesian model to represent ideological stances.
axioms (2)
  • domain assumption Voting can be modeled as a logistic function of latent ideal points plus separate conviction and conformity terms
    Core modeling choice that enables the disentanglement; invoked in the description of the Bayesian logistic regression.
  • domain assumption LLM-inferred topics and stances accurately reflect debate content without systematic bias
    Necessary for applying the model across the 48 topics; no validation details given.

pith-pipeline@v0.9.1-grok · 5732 in / 1464 out tokens · 23054 ms · 2026-06-28T07:47:05.433876+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 14 canonical work pages

  1. [1]

    A Corpus for Modeling User and Language Effects in Argumentation on Online Debating

    Durmus, Esin and Cardie, Claire. A Corpus for Modeling User and Language Effects in Argumentation on Online Debating. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019. doi:10.18653/v1/P19-1057

    work page doi:10.18653/v1/p19-1057 2019

  2. [2]

    Exploring the Role of Prior Beliefs for Argument Persuasion

    Durmus, Esin and Cardie, Claire. Exploring the Role of Prior Beliefs for Argument Persuasion. Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018. doi:10.18653/v1/N18-1094

    work page doi:10.18653/v1/n18-1094 2018

  3. [3]

    Political Analysis , author=

    Birds of the Same Feather Tweet Together: Bayesian Ideal Point Estimation Using Twitter Data , volume=. Political Analysis , author=. 2015 , pages=. doi:10.1093/pan/mpu011 , number=

    work page doi:10.1093/pan/mpu011 2015

  4. [4]

    Jost and Jonathan Nagler and Joshua A

    Pablo Barberá and John T. Jost and Jonathan Nagler and Joshua A. Tucker and Richard Bonneau , title =. Psychological Science , volume =. 2015 , doi =

    2015

  5. [5]

    2003 , publisher =

    Bayesian data analysis , author =. 2003 , publisher =

    2003

  6. [6]

    Stan Modeling Language Users Guide and Reference Manual , year =

  7. [7]

    Current Trends in Bayesian Methodology with Applications , editor =

    Betancourt, Michael and Girolami, Mark , title =. Current Trends in Bayesian Methodology with Applications , editor =. 2015 , pages =

    2015

  8. [8]

    Political Psychology , volume =

    Boutyline, Andrei and Willer, Robb , title =. Political Psychology , volume =. doi:https://doi.org/10.1111/pops.12337 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1111/pops.12337 , abstract =

    work page doi:10.1111/pops.12337

  9. [9]

    Sentence- BERT : Sentence Embeddings using S iamese BERT -Networks

    Reimers, Nils and Gurevych, Iryna. Sentence- BERT : Sentence Embeddings using S iamese BERT -Networks. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019. doi:10.18653/v1/D19-1410

    work page doi:10.18653/v1/d19-1410 2019

  10. [10]

    American Political Science Review , author=

    The Statistical Analysis of Roll Call Data , volume=. American Political Science Review , author=. 2004 , pages=. doi:10.1017/S0003055404001194 , number=

    work page doi:10.1017/s0003055404001194 2004

  11. [11]

    Political Analysis , author=

    Multidimensional Analysis of Roll Call Data via Bayesian Simulation: Identification, Estimation, Inference, and Model Checking , volume=. Political Analysis , author=. 2001 , pages=. doi:10.1093/polana/9.3.227 , number=

    work page doi:10.1093/polana/9.3.227 2001

  12. [12]

    Poole and Howard Rosenthal , journal =

    Keith T. Poole and Howard Rosenthal , journal =. A Spatial Model for Legislative Roll Call Analysis , urldate =

  13. [13]

    , author =

    A study of normative and informational social influences upon individual judgment. , author =. The journal of abnormal and social psychology , volume =. 1955 , publisher =

    1955

  14. [14]

    , title =

    Skitka, Linda J. , title =. Social and Personality Psychology Compass , volume =. doi:https://doi.org/10.1111/j.1751-9004.2010.00254.x , url =. https://compass.onlinelibrary.wiley.com/doi/pdf/10.1111/j.1751-9004.2010.00254.x , year =

    work page doi:10.1111/j.1751-9004.2010.00254.x 2010

  15. [15]

    McCright and Riley E

    Aaron M. McCright and Riley E. Dunlap , title =. The Sociological Quarterly , volume =. 2011 , publisher =. doi:10.1111/j.1533-8525.2011.01198.x , URL =

    work page doi:10.1111/j.1533-8525.2011.01198.x 2011

  16. [16]

    Nature Climate Change , volume =

    Growing polarization around climate change on social media , author =. Nature Climate Change , volume =. 2022 , publisher =

    2022

  17. [17]

    The ANES Guide to Public Opinion and Electoral Behavior , year =

  18. [18]

    The Emotional Dog and Its Rational Tail: A Social Intuitionist Approach to Moral Judgment , booktitle=

    Haidt, Jonathan , editor=. The Emotional Dog and Its Rational Tail: A Social Intuitionist Approach to Moral Judgment , booktitle=. 2008 , pages=

    2008

  19. [19]

    Brandt and Christine Reyna and John R

    Mark J. Brandt and Christine Reyna and John R. Chambers and Jarret T. Crawford and Geoffrey Wetherell , title =. Current Directions in Psychological Science , volume =. 2014 , doi =

    2014

  20. [20]

    Saltzstein , journal =

    Elliot Turiel and Carolyn Hildebrandt and Cecilia Wainryb and Herbert D. Saltzstein , journal =. Judging Social Issues: Difficulties, Inconsistencies, and Consistencies , urldate =

  21. [21]

    Taylor , title =

    Lev Muchnik and Sinan Aral and Sean J. Taylor , title =. Science , volume =. 2013 , doi =

    2013

  22. [22]

    Salganik and Peter Sheridan Dodds and Duncan J

    Matthew J. Salganik and Peter Sheridan Dodds and Duncan J. Watts , title =. Science , volume =. 2006 , doi =

    2006

  23. [23]

    and Goldstein, Noah J

    Cialdini, Robert B. and Goldstein, Noah J. Social Influence: Compliance and Conformity. Annual Review of Psychology. 2004. doi:https://doi.org/10.1146/annurev.psych.55.090902.142015

    work page doi:10.1146/annurev.psych.55.090902.142015 2004

  24. [24]

    Public opinion quarterly , volume =

    Affect, not ideology: A social identity perspective on polarization , author =. Public opinion quarterly , volume =. 2012 , publisher =

    2012

  25. [25]

    , author =

    Heuristic versus systematic information processing and the use of source versus message cues in persuasion. , author =. Journal of personality and social psychology , volume =. 1980 , publisher =

    1980

  26. [26]

    Journal of Political Economy , volume =

    Bikhchandani, Sushil and Hirshleifer, David and Welch, Ivo , title =. Journal of Political Economy , volume =. 1992 , doi =

    1992

  27. [27]

    2009 , publisher =

    Going to extremes: How like minds unite and divide , author =. 2009 , publisher =

    2009

  28. [28]

    Proceedings of the National Academy of Sciences , volume =

    Jan Lorenz and Heiko Rauhut and Frank Schweitzer and Dirk Helbing , title =. Proceedings of the National Academy of Sciences , volume =. 2011 , doi =

    2011

  29. [29]

    and Glance, Natalie , title =

    Adamic, Lada A. and Glance, Natalie , title =. Proceedings of the 3rd International Workshop on Link Discovery , pages =. 2005 , isbn =. doi:10.1145/1134271.1134277 , abstract =

    work page doi:10.1145/1134271.1134277 2005

  30. [30]

    Nature Human Behaviour , volume =

    Political polarization of news media and influencers on Twitter in the 2016 and 2020 US presidential elections , author =. Nature Human Behaviour , volume =. 2023 , publisher =

    2016

  31. [31]

    Proceedings of the National Academy of Sciences , volume =

    Petter Törnberg , title =. Proceedings of the National Academy of Sciences , volume =. 2022 , doi =

    2022

  32. [32]

    and Hanson, Brittany E

    Skitka, Linda J. and Hanson, Brittany E. and Morgan, G. Scott and Wisneski, Daniel C. The Psychology of Moral Conviction. Annual Review of Psychology. 2021. doi:https://doi.org/10.1146/annurev-psych-063020-030612

    work page doi:10.1146/annurev-psych-063020-030612 2021

  33. [33]

    1984 , publisher =

    The spatial theory of voting: An introduction , author =. 1984 , publisher =

    1984

  34. [34]

    2004 , publisher =

    Item response theory: Parameter estimation techniques , author =. 2004 , publisher =

    2004

  35. [35]

    Aramovich and Brad L

    Nicholas P. Aramovich and Brad L. Lytle and Linda J. Skitka , title =. Social Influence , volume =. 2012 , publisher =. doi:10.1080/15534510.2011.640199 , URL =

    work page doi:10.1080/15534510.2011.640199 2012

  36. [36]

    Journal of the American Statistical Association , year =

    Luca Alessandro Silva and Giacomo Zanella , title =. Journal of the American Statistical Association , volume =. 2024 , publisher =. doi:10.1080/01621459.2023.2257893 , URL =

    work page doi:10.1080/01621459.2023.2257893 2024

  37. [37]

    Nature Human Behaviour , volume =

    A semantic embedding space based on large language models for modelling human beliefs , author =. Nature Human Behaviour , volume =. 2025 , publisher =

    2025

  38. [38]

    arXiv preprint arXiv:2603.11253 , year =

    LLMs Can Infer Political Alignment from Online Conversations , author =. arXiv preprint arXiv:2603.11253 , year =

    arXiv

  39. [39]

    A Methodological Guide on Using Large Language Models for Text Annotation in the Social Sciences and Humanities with Python and R , doi =

    Fang, Qixiang and Garcia-Bernardo, Javier and Kesteren, Erik-Jan , year =. A Methodological Guide on Using Large Language Models for Text Annotation in the Social Sciences and Humanities with Python and R , doi =

  40. [40]

    2026 , eprint=

    Emergence of Stereotypes and Affective Polarization from Belief Network Dynamics , author=. 2026 , eprint=

    2026

  41. [41]

    Proceedings of the National Academy of Sciences , volume =

    Matteo Cinelli and Gianmarco De Francisci Morales and Alessandro Galeazzi and Walter Quattrociocchi and Michele Starnini , title =. Proceedings of the National Academy of Sciences , volume =. 2021 , doi =

    2021