Addressing the challenges of utilising tracking-based smartphone travel survey data in travel demand modelling

Tracking-based smartphone travel surveys are increasingly used to collect travel behaviour data semi-passively, where inference algorithms identify key attributes such as travel modes and trip purposes, and participants are asked to verify or correct them. However, verification rates are often low, resulting in substantial amounts of untagged or unlabelled data. Researchers typically either discard such observations, thereby loose valuable information, or utilise the deterministically inferred labels assigned from inference algorithms in subsequent analysis. However, it remains to be evaluated whether such inferred untagged data can be reliably used to develop travel demand models.

This thesis addresses this research gap by examining the benefits and implications of using untagged datasets to develop travel demand models, and by proposing methods to more accurately infer missing trip information, particularly trip purpose. Three empirical studies are conducted: Study I evaluates the impact of untagged trip purposes in a daily trip generation model; Study II exploits the panel nature of the collected data to better impute missing trip purposes and assesses the impact of utilising untagged trip purposes in a mode choice model; and Study III evaluates estimation approaches for mode choice models developed using untagged data. These cases reflect uncertainty in either the dependent variable (Studies I and III) or a discrete independent variable (Study II).

Results show that the benefits of utilising untagged data depend not only on inference accuracy, but also on which variable is inferred. In our case studies, when the dependent variable is inferred (which leads to larger sample sizes), the parameter estimates become more precise (lower standard errors) but not necessarily more accurate (closer to population-level parameters). In contrast, when an independent variable is inferred, which implies an additional explanatory variable is available in the estimated model, the parameters also become more accurate.

The thesis also demonstrates that it is better to utilise untagged variables while accounting for uncertainty in the inference process rather than considering deterministic labels. Additionally, exploiting the panel structure of the untagged datasets by making use of posterior analysis in a mixed multinomial logit model improves the imputation of missing trip purposes compared to standard approaches.

Overall, the thesis highlights both opportunities and limitations of using untagged travel data and provides methodological guidance for incorporating such data into travel demand models.