How Can You Think That? Deliberation and the Learning of Opposing Arguments

Gaurav Sood, Robert C. Luskin, and James S. Fishkin

Paper · Research design · Data

Question and motivation

Does deliberation help people articulate reasons for policy views they oppose? Learning what another side can argue is different from changing one's own position. The study examines whether a Deliberative Poll on school policy in Omagh, Northern Ireland, broadens the reasons people can give on both sides.

Data and research design

The analysis combines a numeric survey, discussion-group roster, and coder labels for open-ended answers. It compares participants with controls at the follow-up survey and follows participants from the event to follow-up. Event attendance was self-selected, and the open-ended outcome was not measured at recruitment. The participant–control comparison therefore estimates an association, not a causal effect of attendance.

Key findings

At follow-up, participants give more coded reasons on average than controls, but the interval for the adjusted difference includes zero. Participants' reason counts decline between the event and follow-up, also with an interval spanning zero. The data do not clearly establish a lasting participant advantage.

Comparison	Difference [95% interval]	Respondents
Follow-up participants minus controls	1.12 [-0.68, 2.92]	185
Adjusted participant–control difference	0.97 [-0.83, 2.77]	174
Follow-up minus event among participants	-0.87 [-3.44, 1.70]	62

Estimates are differences in the number of answers containing a coded reason; intervals are 95% confidence intervals. Uncertainty uses CR2 small-sample corrections, grouping participants by discussion group and treating controls as singleton clusters. The noncontemporaneous event–control comparison is reported in tabs/main_estimates.csv, alongside these comparisons.

Reproduce

The project uses R and XeLaTeX. From the repository root:

git clone --branch v0.4.5 --depth 1 https://github.com/soodoku/dp_data.git ../dp_data
make restore
make check

The public dp_data release supplies all three analytical inputs. A downloaded release archive works too; extract it to ../dp_data, or set DP_DATA_ROOT to its location. Builds verify the pinned file hashes before reading data. See docs/data.md for the source contracts and update procedure.

The individual targets are:

make analysis
make lint
make test
make paper
make ci-docker

The compiled paper and Supporting Information are in ms/main.pdf. Analytical results are stored together in the generated data/derived/analysis_results.rds object. The pipeline renders publication tables directly with knitr::kable; the two headline CSV files in tabs/ provide machine-readable estimates. Inline manuscript numbers come from tabs/numbers.tex.

Edit docs/README.in.md to change this README; scripts/05_exhibits.R supplies the current result table.

Files and pipeline

• data/: source manifest and ignored generated analysis data
• scripts/00_config.R: paths, data contracts, and analysis settings
• scripts/00_utils.R: reusable functions
• scripts/01_prepare_survey.R and 02_prepare_coding.R: input preparation
• scripts/03_build_analysis.R and 04_estimate.R: analysis data and estimates
• scripts/05_exhibits.R: manuscript exhibits, inline numbers, and this README
• scripts/98_validate.R: analytical validation
• scripts/99_run_all.R: run the stages in order
• tests/: data-contract and coding tests
• tabs/: headline CSV results and generated LaTeX table fragments
• figs/: generated publication figures
• ms/: manuscript, bibliography, and compiled paper
• docs/: research-design and data documentation

The source files used by the current pipeline and their SHA-256 hashes are pinned in data/manifest.yaml. Sample membership, demographics, policy attitudes, and group identifiers are reconstructed from the raw survey export and group roster; historical derived survey variables are not used.

Main measurement decisions

The main outcome counts answers that give at least one reason for a school-policy view. Answers marked nonsense, missing, or vague do not count. For three questions, code 3 means the answer calls people with that view prejudiced or ignorant. The main count leaves out that code; a separate check includes it. Another valid code in the same answer still counts. When coders disagree and no third coder settles it, the main count is missing for that respondent and survey. We also report results from each original coder separately.

Non-administered survey waves remain missing. They are never converted to zero. Uncertainty estimates use CR2 small-sample corrections, clustering participants by discussion group while treating controls as singleton clusters.

Data responsibility

The public inputs contain a numeric survey, group roster, and coder labels. Verbatim responses are excluded. Historical files containing response text remain in the original repository history and the local source archive; they are not required to reproduce the paper. See docs/data.md.

The MIT license covers code and documentation only. It does not relicense the source data.

🔗 Adjacent Repositories

• soodoku/dp_misinformation — Deliberation and Misinformation (Luskin and Sood). Data, code, and manuscript.
• soodoku/dp_learning — Deliberation and Learning: Evidence From Deliberative Polls (Luskin, Sood, Fishkin)
• soodoku/dp_data — Deliberative Poll data
• soodoku/dp_distortions — Replication Data and Scripts for Deliberative Distortions
• soodoku/sophisticated_cynic — Is Cynicism Taken as Evidence of Political Sophistication? No.

Powered by Adjacent