Robert C. Luskin, Gaurav Sood, Yul Min Park, and Joshua Blank
How much apparent misinformation comes from the way survey questions are asked? Media-poll items often invite opinions and omit an explicit “don't know” option. Those features can turn guesses into answers that headlines present as false beliefs. The companion paper, The Waters of Casablanca, develops the distinction between misinformation and less confident belief and the 0–10 measure used here.
The study codes the wording and response options of media-poll misinformation items, then tests question design in a randomized survey experiment. Mechanical Turk respondents received one of five versions of nine items: media wording, versions changing don't-know options and claim cues, or a confidence scale. The corpus describes polling practice; random assignment identifies the effects of the tested question versions within the experimental sample. The sample is not representative of the U.S. public.
Offering and encouraging “don't know” substantially reduces incorrect answers. A confidence scale finds still less misinformation. Adding background and repeating the false claim has little average effect in the tested comparisons. Incorrect multiple-choice answers therefore overstate confidently held false beliefs in these experiments.
Points show the share answering each item incorrectly, with 95% Wilson intervals. For the scale, an incorrect answer means choosing the false endpoint. Don't-know and skipped answers count as not incorrect.
make restore
make check
make restore installs the package versions in renv.lock. make check rebuilds the analysis and manuscript, lints the R code, and runs the tests. R 4.6 and XeLaTeX with latexmk are required. make ci-docker runs these checks in the project's standard Rocker image.
| Path | Contents |
|---|---|
data/raw/ |
The media-poll corpus, Roper toplines, and the survey experiment; see data/README.md |
docs/ |
Item-level decisions, answer keys, the questionnaire, and how each citation was checked |
R/ |
Reusable functions: source validation, corpus scoring, experimental estimates, and output writers. These files do not run the analysis. |
scripts/ |
Configuration and numbered execution stages, listed below. |
data/derived/ |
Generated intermediate data; ignored by Git. |
ms/ |
main.tex, references.bib, and the compiled main.pdf |
tests/testthat/ |
Reproductions of the 2018 draft's numbers, tests of the answer keys, and a check that the survey data carry no direct identifiers |
make analysis runs scripts/99_run_all.R. It loads the configuration and reusable functions in R/, then executes stages 01–04 in separate environments. Stages pass results through files.
| File | Purpose |
|---|---|
scripts/00_config.R |
Paths, source hashes, labels, colors, theme_paper(), figure sizes, and table_style. |
scripts/01_prepare_data.R |
Verify source hashes, read and score the inputs, and save data/derived/prepared_data.rds. |
scripts/02_estimate.R |
Read the prepared data and write estimates to tabs/*.csv. |
scripts/03_figures.R |
Read the estimates and write PDF and PNG figures to figs/. |
scripts/04_tables.R |
Generate LaTeX tables, manuscript numbers, and tabs/style.tex. |
scripts/99_run_all.R |
Run all stages in order. |
Change plot and table defaults in 00_config.R. The manuscript applies the generated \TableStyle to each table. Statistical and coding functions stay in R/corpus.R and R/experiment.R, where tests can call them without running the pipeline.
- finite-sample/know_casablanca — The Waters of Casablanca: On Political Misinformation (Luskin and Sood). Data, code, and manuscript.
- finite-sample/know_hidden — Mis-measuring Political Knowledge? Do People Know More—or Even Less—about Politics than Commonly Thought?
- finite-sample/stddev — Is the public's fickleness the work of a few?
- finite-sample/endorsement — Derive Proportion Supporting A Group From an Endorsement Experiment
- finite-sample/know_pgap_cues — Partisan cues and evaluations of the same economic information (Roush and Sood)
Powered by Adjacent