# Extraction report: shurakov2023trainflightretraction

## Summary
- Created a separate unpublished-study package for Shurakov's removed Experiment 3 rather than modifying `papers/shurakovndstakeseffectnew`.
- Source draft: `source/RetractionBased_PaperDraft_18_11_N.docx`; converted PDF: `pdf/paper.pdf`; extracted text: `out/fulltext.md` and `out/text.txt`.
- Source email: `source/email.eml`; it states that the published Shurakov paper originally included an additional Experiment 3 with Train and Flight scenarios.
- Primary computation file: `analysis/effect_sizes.qmd`.

## Extraction Decisions
- Extracted two primary meta-analytic effects:
  - `s1_e1`: Train, Stakes vs Neutral.
  - `s1_e2`: Flight, Stakes vs Neutral.
- Did not extract Evidence vs Neutral or Evidence vs Stakes as primary effects, because Evidence is a defeater condition rather than the low-vs-high stakes contrast.
- Used the composite retraction score: binary response (`I do` = +1, `I don't` = -1) times confidence (1-7), following the draft and author Python script.
- Used `esc::esc_mean_sd` for SMDs with project sign convention `d = mean(low) - mean(high)`.
- Moderator coding update: after clarifying that `awareness`, `evidence`, and `evidence_reliability` are coded from the subject
  of the target knowledge attribution, both Train (`s1_e1`) and Flight (`s1_e2`) are coded `awareness: No`. The target
  subjects are the guy/woman, not the participant-attributor; their evidence codings remain `External` for Train and
  `First Person` for Flight.

## Data Handling
- The author sent three CSV exports:
  - `data/Finished_Responses.csv`
  - `data/Unfinished_Responses.csv`
  - `data/Retraction_Third_person_April_14_2023.csv`
- The timestamped `Retraction_Third_person_April_14_2023.csv` contains finished rows only and does not reproduce the manuscript's Train/Flight Neutral and Stakes Ns as well as the combined Finished + Unfinished exports.
- The QMD therefore combines `Finished_Responses.csv` and `Unfinished_Responses.csv`, then filters to consenting rows with initial knowledge endorsement, correct attention check, and usable follow-up response.

## Issues For Review

### HUMAN_CHECK_SHURAKOV_EXP3_N

The draft reports 361 recruited participants, 181 Train, 180 Flight, no attention-check failures, 20 Train sceptics, and 4 Flight sceptics. Combining the supplied Finished and Unfinished exports gives 365 consenting rows, 181 Train-assigned rows, 182 Flight-assigned rows, 20 Train sceptics, 4 Flight sceptics, and 338 usable composite-score responses across Train/Flight conditions. The focal Neutral/Stakes means and Ns match the draft closely, but the global participant counts do not align perfectly.

Current decision: compute effect sizes from the combined author-sent CSVs and flag `manuscript_dataset_n_discrepancy`.

### HUMAN_CHECK_SHURAKOV_EXP3_FLIGHT_EVIDENCE

The Flight Evidence group is not extracted as a primary stakes effect, but it is useful for checking manuscript consistency. The combined CSVs yield Flight Evidence `n = 59`, whereas the draft's reported omnibus df and pairwise Ns imply `n = 58`. This does not affect the extracted Flight Stakes-vs-Neutral effect, where both groups are `n = 59`.

## Validation
- YAML schema validation passed against `docs/stakes_meta_schema.json`.
- Quarto computation rendered: `analysis/effect_sizes.html`.
- Extraction HTML rendered: `shurakov2023trainflightretraction.html`.
- Effects export rebuilt: `out/effects_with_unpublished.csv` now contains 312 rows, including the two Shurakov unpublished effects.