How Effective are Marker Variables at Predicting Attitudinal Factor Scores? An Out-of-sample Evaluation

Details

• Creators:
  Soria, Jason ; Mokhtarian, Patricia L.
• Corporate Creators:
  Center for Understanding Future of Travel Behavior and Demand (TBD) National University Transportation Center (UTC) ; Georgia Institute of Technology. School of Civil and Environmental Engineering
• Corporate Contributors:
  United States. Department of Transportation. University Transportation Centers (UTC) Program ; United States. Department of Transportation. Office of the Assistant Secretary for Research and Technology
• Subject/TRT Terms:
  Attitudes Data Files Machine Learning Predictive Models Surveys Travel Behavior Travel Demand
• Resource Type:
  Tech Report
• Geographical Coverage:
  United States
• Edition:
  Final Report, 2023-2024
• Corporate Publisher:
  Center for Understanding Future of Travel Behavior and Demand (TBD) National University Transportation Center (UTC)
• Abstract:
  This report is one in a series of studies designed to investigate the practicality of including attitudes as explanatory variables in practice-oriented travel demand forecasting models. In this study, we respectively applied random forest (RF) and elastic net regression (ENR) to 15 marker variables (MVs), to predict factor scores on four attitudes: pro-car ownership, pro-non-car alternatives, pro-suburban, and urbanite. We incorporated those four imputed attitudes into multinomial logit vehicle ownership (VO) models and compared the results to those of models including the original four factor scores, models containing only the four MVs most strongly associated with the same four attitudes, and models containing no attitudes. We created 1,000 random splits of a sample of 3,178 responses to a 2017 survey of Georgia adults, training the RF and ENR functions on each donor half-sample and applying those functions to the recipient half-sample. We reported results averaged over the 1,000 recipient half-samples. In the VO models, estimated coefficients for all sets of attitudes (original, ENR-imputed, RF-imputed, and MVs only) were by far most often both statistically significant and with the expected sign. Perhaps most importantly, the predictive power of the models markedly improved specifically for zero-car households whenever the attitudes were included. Using only the marker variables themselves gave results nearly as good as those associated with the more elaborate prediction of factor scores using machine learning methods.
• Format:
  PDF
• Funding:
  69A3552344815, 69A3552348320
• Collection(s):
  University Transportation Centers
• Main Document Checksum:
  urn:sha-512:88251496225af1ecccea454d48eca20bf54b0cfb42e8ab959e9554a4561bd1cea8c18ea9ed28971e11420c9d2ed027559c28d3ce5d8e5fae0c62ebd6b4aa3427
• Download URL:
  https://rosap.ntl.bts.gov/view/dot/87066/dot_87066_DS1.pdf
• File Type:
  [PDF - 1.95 MB ]