Cohort cleaning & RQ-aligned EDA

Analytic sample support for primary and secondary research questions

Notebook: notebooks/cleaning_eda_full_cohort.ipynb
Code: src/ca_personas/load.py · eda.py
CLI: ca-personas prepare --join inner
Artifacts: outputs/eda/
Merge sanity check: merge_coverage_sanity.md

1 Why this page exists

Every persona tier, ML baseline, and secondary RQ rests on the same cleaned analytic sample. This page records how that sample is built and what the empirical CA patterns look like by employment, transit, sex, student status, and country — the slices used later when interpreting LLM stereotyping error. The demos persona tier’s base demographics layer is Age, Sex, Country of residence, and Student status.

2 Sample construction

Step N
Prolific File A + File B (unique IDs) 262
Qualtrics File C rows 273
Matched Prolific ∩ Qualtrics 252
Qualtrics-only (disregard) 21
Prolific-only (disregard) 10
Dropped incomplete PRCA items 11
Analytic sample (complete group + interpersonal GT) 241

Waves (matched respondents): File A = 99, File B = 153 of the 252 Prolific∩Qualtrics matches. Full Prolific waves omit Ethnicity / Nationality / Language; the demos tier therefore uses Age, Sex, Country of residence, and Student status.

Covariate coverage in the analytic sample: employment 100%, transit (Q26–Q29) 100%, Student status 93.0% non-missing (17/241 DATA_EXPIRED → missing).

3 Ground-truth CA

Subscale n Mean SD Median Range
Group CA 241 14.62 6.03 13 6–30
Interpersonal CA 241 14.31 5.82 13 6–30

Band prevalence (low ≤13 / moderate 14–19 / high ≥20): group 53.9% / 25.3% / 20.7%; interpersonal 53.1% / 28.6% / 18.3%.

4 RQ1 lens — employment

Mean CA by employment
Employment n Mean group CA Mean interpersonal CA % high group
Full-Time 148 13.32 12.99 12.8%
Part-Time 31 16.16 16.10 25.8%
Other 62 16.95 16.56 37.1%

Full-time respondents report lower CA. If an LLM stereotypes “unemployed / other → higher anxiety,” that pattern has some empirical footing in this sample — residual error analyses should still check whether the model over-applies it.

5 RQ2 lens — public transit (Q26)

Q26 n Mean group CA Mean interpersonal CA
Never 50 17.38 16.34
0–1 days/month 37 15.38 15.16
2–4 days/month 53 14.51 13.72
4–8 days/month 46 13.28 13.48
8+ days/month 55 12.84 13.16

Higher ridership ↔︎ lower mean CA (Never group M = 17.38 → 8+ days M = 12.84). This is the descriptive foundation for Transit → CA (memo) and CA → Transit (memo).

6 Stereotyping lens — sex, student status & country

Slice n Mean group CA Mean interpersonal CA
Female 120 14.88 14.75
Male 121 14.37 13.88
United States 158 14.04 13.90
United Kingdom 56 16.34 15.54
Canada 20 14.65 15.15
Student = Yes 34 16.35 17.03
Student = No 190 14.26 13.77

Sex differences are small. UK means sit higher than US in this sample; students show elevated interpersonal CA. Sparse countries (n = 1–2) are not interpretable alone.

7 Implication for the manuscript

  • Persona tiers that add employment or transit are adding covariates that track true CA in the cohort.
  • Bias analyses should compare LLM absolute error against these empirical group means — not against a flat “everyone is average” null.
  • Secondary RQs convert the transit gradient into formal tests and predictive metrics.

8 Reproducibility

ca-personas prepare --join inner
jupyter nbconvert --to notebook --execute notebooks/cleaning_eda_full_cohort.ipynb