Do car access, employment, or ride-share predict regular transit?

Secondary research questions — geo-memo follow-up Random Forests

Project: PSYCH 755 — CA persona / PRCA framework
Analysis code: src/ca_personas/transit_covariate_rf.py
Notebooks:
secondary_rq_car_access_transit_rf.ipynb ·
secondary_rq_employment_transit_rf.ipynb ·
secondary_rq_rideshare_transit_rf.ipynb ·
secondary_rq_transit_covariate_followups.ipynb
CLI: ca-personas covariate-transit-rf --join inner --seed 42
Artifacts: outputs/transit_covariate_rf/
Memos: memos/transit_covariate_followups.qmd and family memos under memos/

1 Research questions

The geography → transit memo (memos/geo_predicts_transit.qmd) asked which other features might predict regular public-transit use after finding lat/long discrimination of AUC = 0.551. Communication-apprehension scores were already tested (AUC = 0.590; memos/ca_scores_predict_transit.qmd). This write-up evaluates the remaining named candidates:

Driver’s license & car access (Q20, Q21)
Employment status
Ride-share frequency (Q28, Q29)
Joint mobility bundle (all of the above)

2 Methods

2.1 Sample & outcome

Sources: Prolific File A + File B stacked, inner-joined to Qualtrics File C
Base analytic cohort: n = 241 with complete PRCA group/interpersonal items and usable Q26
Outcome: regular_transit = Q26 ∈ {4-8 days a month, 8 or more days a month}
Complete-case frames: car access n = 149; employment n = 241; ride-share n = 233; bundle n = 143

2.2 Models

For each feature family: OneHotEncoder → balanced RandomForestClassifier (500 trees, min_samples_leaf=3), stratified 5-fold cross_val_predict probabilities, Gini + permutation importance (scoring="roc_auc", 30 repeats). Seed = 42.

Benchmarks pasted from companion seeded runs: geo AUC = 0.551; CA AUC = 0.590; chance = 0.500.

3 Results

Spec	n	ROC-AUC	AP	Bal. acc.	F1
Q28 only	241	0.762	0.689	0.730	0.702
Mobility bundle	143	0.746	0.604	0.717	0.667
Ride-share (Q28/Q29)	233	0.745	0.655	0.731	0.709
Car access (Q20/Q21)	149	0.607	0.515	0.639	0.476
CA benchmark	241	0.590	—	—	—
Q27 only	239	0.589	0.513	0.623	0.467
Geo benchmark	241	0.551	—	—	—
Employment	241	0.528	0.424	0.560	0.586
Chance	—	0.500	—	0.500	—

Ride-share days (Q28) dominate: regular-transit prevalence rises from 15.7% (Never) to 94.1% (8+ days/month). Car access (Q21) is the main car-family driver (No access = 76.0% regular vs Yes = 30.9%). Employment differences are small.

4 Interpretation

Mobility self-reports—especially ride-share frequency—predict weekly+ transit far better than survey geolocation or CA scores in this cohort. The joint bundle’s near-parity with ride-share alone suggests limited incremental value from car/employment once Q28 is included, on the smaller overlapping sample. Associations are descriptive, not causal.

5 Limitations

Unequal complete-case N across families; car items are ~38% missing.
Same-wave survey items; shared method variance between Q26 and Q28 is possible.
Coarse employment coding.
Qualtrics/Prolific convenience sample; not population transit forecasting.

--- title: "Do car access, employment, or ride-share predict regular transit?" subtitle: "Secondary research questions — geo-memo follow-up Random Forests" --- **Project:** PSYCH 755 — CA persona / PRCA framework **Analysis code:** [`src/ca_personas/transit_covariate_rf.py`](../src/ca_personas/transit_covariate_rf.py) **Notebooks:** [`secondary_rq_car_access_transit_rf.ipynb`](../notebooks/secondary_rq_car_access_transit_rf.ipynb) · [`secondary_rq_employment_transit_rf.ipynb`](../notebooks/secondary_rq_employment_transit_rf.ipynb) · [`secondary_rq_rideshare_transit_rf.ipynb`](../notebooks/secondary_rq_rideshare_transit_rf.ipynb) · [`secondary_rq_transit_covariate_followups.ipynb`](../notebooks/secondary_rq_transit_covariate_followups.ipynb) **CLI:** `ca-personas covariate-transit-rf --join inner --seed 42` **Artifacts:** `outputs/transit_covariate_rf/` **Memos:** [`memos/transit_covariate_followups.qmd`](../memos/transit_covariate_followups.qmd) and family memos under `memos/` --- ## Research questions The geography → transit memo ([`memos/geo_predicts_transit.qmd`](../memos/geo_predicts_transit.qmd)) asked which other features might predict regular public-transit use after finding lat/long discrimination of AUC = **0.551**. Communication-apprehension scores were already tested (AUC = **0.590**; [`memos/ca_scores_predict_transit.qmd`](../memos/ca_scores_predict_transit.qmd)). This write-up evaluates the remaining named candidates: 1. Driver's license & car access (`Q20`, `Q21`) 2. Employment status 3. Ride-share frequency (`Q28`, `Q29`) 4. Joint mobility bundle (all of the above) ## Methods ### Sample & outcome - **Sources:** Prolific File A + File B stacked, inner-joined to Qualtrics File C - **Base analytic cohort:** n = 241 with complete PRCA group/interpersonal items and usable `Q26` - **Outcome:** `regular_transit` = `Q26` ∈ {`4-8 days a month`, `8 or more days a month`} - **Complete-case frames:** car access n = 149; employment n = 241; ride-share n = 233; bundle n = 143 ### Models For each feature family: `OneHotEncoder` → balanced `RandomForestClassifier` (500 trees, `min_samples_leaf=3`), stratified 5-fold `cross_val_predict` probabilities, Gini + permutation importance (`scoring="roc_auc"`, 30 repeats). Seed = 42. Benchmarks pasted from companion seeded runs: geo AUC = 0.551; CA AUC = 0.590; chance = 0.500. ## Results | Spec | n | ROC-AUC | AP | Bal. acc. | F1 | |---|---:|---:|---:|---:|---:| | **Q28 only** | 241 | **0.762** | 0.689 | 0.730 | 0.702 | | Mobility bundle | 143 | 0.746 | 0.604 | 0.717 | 0.667 | | Ride-share (Q28/Q29) | 233 | 0.745 | 0.655 | 0.731 | 0.709 | | Car access (Q20/Q21) | 149 | 0.607 | 0.515 | 0.639 | 0.476 | | CA benchmark | 241 | 0.590 | — | — | — | | Q27 only | 239 | 0.589 | 0.513 | 0.623 | 0.467 | | Geo benchmark | 241 | 0.551 | — | — | — | | Employment | 241 | 0.528 | 0.424 | 0.560 | 0.586 | | Chance | — | 0.500 | — | 0.500 | — | **Ride-share days (`Q28`)** dominate: regular-transit prevalence rises from 15.7% (Never) to 94.1% (8+ days/month). **Car access (`Q21`)** is the main car-family driver (No access = 76.0% regular vs Yes = 30.9%). **Employment** differences are small. ## Interpretation Mobility self-reports—especially ride-share frequency—predict weekly+ transit far better than survey geolocation or CA scores in this cohort. The joint bundle’s near-parity with ride-share alone suggests limited incremental value from car/employment once `Q28` is included, on the smaller overlapping sample. Associations are descriptive, not causal. ## Limitations - Unequal complete-case *N* across families; car items are ~38% missing. - Same-wave survey items; shared method variance between `Q26` and `Q28` is possible. - Coarse employment coding. - Qualtrics/Prolific convenience sample; not population transit forecasting.