.png){kind=icon}
.png){kind=icon}
Return
Using Multiple Biased Data Sets to Recover Missing Trips with a Behaviorally Informed Model
.png){kind=icon}
PRE
AI
DOI：10.1287/trsc.2024.0550.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Trip generation, a critical first step in travel demand forecasting, requires not only estimating trips from the observed sample data, but also calculating the total number of trips in the population, including both the observed trips and the trips missed from the sample (we call them missing trips in this paper). The latter, how to recover missing trips, is scarcely studied in the academic literature, and the state-of-the-art practice is through the application of sample weights to extrapolate from observed trips to the population total. In recent years, big location-based service (LBS) has become a promising alternative data source (in addition to household travel survey data) in trip generation. Because users self-select into using different mobile services that result in LBS data, selection bias exists in the LBS data, and the kinds of trips excluded or included differ systematically among data sources. This study addresses this issue and develops a behaviorally informed approach to quantify the selection biases and recover missing trips. The key idea is that because biases reflected in different data sources are likely different, the integration of multiple biased data sources will mitigate biases. This is achieved by formulating a capture probability that specifies the probability of capturing a trip in a data set as a function of various behavioral factors (e.g., socio-demographics and area-related factors) and estimating the associated parameters through maximum likelihood or Bayesian methods. This approach is evaluated through experimental studies that test the effects of data and model uncertainty on its ability of recovering missing trips. The model is also applied to two real-world case studies: one using the 2017 National Household Travel Survey data and the other using two LBS data sets. Our results demonstrate the robustness of the model in recovering missing trips, even when the analyst completely mis-specifies the underlying trip generation process and the capture probability functions (for quantifying selection biases). The developed methodology can be scalable to any number of data sets and is applicable to both big and small data sets.
Keywords:
self-selection bias
location-based service (LBS) data
small and big data
likelihood-based method
travel demand
Bayesian estimation