Memo: Does country of residence predict regular transit?

Research memorandum — dedicated country RF vs lat/long geography

Author

Jack J. Burleson

Published

Invalid Date

Research question: Does country of residence alone predict regular public-transit use, and how does it compare to Qualtrics lat/long geography?

Parent memo: geo_predicts_transit.qmd · Companion: country_car_predicts_transit.qmd

1 Answer, Response, + Summary of Results

On the full matched analytic cohort (n = 241), a categorical Random Forest ((Breiman, 2001)) on Country of residence yields CV ROC-AUC ≈ 0.552—essentially identical to the lat/long geo memo (≈ 0.551) and the geo module’s country-only baseline (≈ 0.549).

Short answer: Country is a weak place cue, interchangeable with survey coordinates for this outcome. It does not unlock the discrimination that mobility items (especially Q28) provide. Figure Figure 1 shows country prevalence and the ROC curve.

Figure 1.. Country prevalence and ROC

1.1 Prevalence by country (n ≥ 20)

Table Table 1 reports prevalence by country.

Table 1.. Country of residence predicting regular transit.
Country n % regular
United States 158 37.3%
United Kingdom 56 48.2%
Canada 20 60.0%

Remaining countries are singleton / tiny cells and should not be over-interpreted.

1.2 Performance

Table Table 2 reports the CV performance.

Table 2.. Model / n / ROC-AUC.
Model n ROC-AUC
Country of residence 241 0.552
Lat/long (geo memo) 241 0.551
CA benchmark 241 0.590
Chance — 0.500

1.3 Interpretation

  1. Answers the geo memo’s open question: ****country vs raw coordinates are tied****, both barely above chance.
  2. Place labels alone cannot explain the stronger rideshare/car associations.
  3. Pairing country with car access does help—see country_car_predicts_transit.qmd.

Sources: ca-personas followup-experiments --experiments country · outputs/followup_experiments/country/ · formal write-up docs/secondary_rq_followup_experiments.md

2 What questions or uncertainties remain?

Would city-level or transit-infrastructure GIS features beat both country and IP lat/long? VPN/IP misplacement concerns from the geo memo still apply to the coordinate arm.

References

Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324