Memo: Do car license & access predict regular transit?

Research memorandum — Q20/Q21 follow-up to the geography memo

Author

Jack J. Burleson

Published

Invalid Date

Research question: Do driver’s license (Q20) and car access (Q21) predict whether a matched respondent takes public transportation regularly?

Parent memo: geo_predicts_transit.qmd
Comparison memo: transit_covariate_followups.qmd

1 Answer, Response, + Summary of Results

Using the Prolific↔︎Qualtrics matched cohort (File A + File B stacked joined to File C; analytic PRCA sample ((McCroskey, 1970); n = 241), we tested the geo-memo follow-up candidate car license / car access. Regular transit is Q26 ∈ {4–8 days a month, 8 or more days a month}. Because Q20/Q21 are missing for many respondents, the complete-case modeling frame is n = 149 (58 regular / 91 not regular; prevalence = 38.9%). A balanced Random Forest ((Breiman, 2001)) with stratified 5-fold CV (seed=42) used Q20 and Q21 as one-hot features.

Short answer: Yes — car license/access exceeds geography and CA on CV ROC-AUC. Car license/access recovers CV ROC-AUC = 0.607, above chance (0.500), the geo RF benchmark (0.551), and the CA RF benchmark (0.590). Access (Q21) dominates license (Q20). Figure Figure 1 shows prevalence by level and the ROC curve.

Figure 1.. Car access prevalence by level and ROC curve

Descriptive associations (complete cases). Respondents without car access ride regularly far more often than those with access (Table Table 1):

Table 1.. Random Forest performance for car access predicting regular transit.
Item Level n % regular
Q21 access No 25 76.0%
Q21 access Yes 123 30.9%
Q20 license No 20 55.0%
Q20 license Yes 129 36.4%

Random Forest (stratified CV). Table Table 2 reports the CV ROC-AUCs.

Table 2.. Model / n / ROC-AUC.
Model n ROC-AUC
Q20 + Q21 RF 149 0.607
CA RF benchmark 241 0.590
Geo RF benchmark 241 0.551
Chance — 0.500

Permutation importance ranks Q21 (mean \(\Delta\)AUC = 0.182) above Q20 (0.014). At a 0.5 threshold the forest is conservative on the positive class (precision = 0.77, recall = 0.34, F1 = 0.48), consistent with a sparse “no access” positive signal.

Conclusion. Car access recovers AUC = 0.607 on complete cases (n = 149): lacking a usable car is associated with 76.0% regular ridership vs 30.9% among those with access, yet item missingness truncates the analytic N. Interpret as a mobility constraint correlate, not a causal estimate.

Sources: notebooks/secondary_rq_car_access_transit_rf.ipynb · src/ca_personas/transit_covariate_rf.py · ca-personas covariate-transit-rf --specs car_access · github.com/Exios66/psych755-jjb · formal write-up docs/secondary_rq_transit_covariate_followups.md

2 What questions or uncertainties remain?

Are Q20/Q21 missing systematically (e.g., skip logic, survey fatigue), and would imputed or survey-design-aware models change the AUC? Does car access mediate part of the CA–transit or geo–transit associations?

3 What other features may also well-predict regular public transit use?

Ride-share frequency (Q28/Q29) recovers AUC = 0.745 (rideshare_predicts_transit.qmd). Employment status alone recovers AUC = 0.528 (employment_predicts_transit.qmd). Head-to-head: transit_covariate_followups.qmd.

References

Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324
McCroskey, J. C. (1970). Measures of communication-bound anxiety. Speech Monographs, 37(4), 269–277. https://doi.org/10.1080/03637757009375677