A Spatio-temporal Approach For Identifying The Sample Size For Gps-based Travel Surveys.

GPS-based travel surveys hold many attractive features that make such surveys very attractive for travel behaviour studies. However, the sample size calculation procedure for such surveys is more complicated compared to well-known and widely applied conventional household one/two-day travel surveys. The higher cost of GPS surveys requires scrutiny at the sample size planning stage to ensure cost effectiveness. Different machine learning-based techniques have been developed to infer the transportation mode based upon speed calculated from GPS data. However, none of these studies calculate the sample size required for validating these techniques. Calculating the most effective sample size for this inference mainly depends on the "variability" of the variables that a survey is designed to investigate. Hence, we use speed as an example of these observed variables. Since different surveys target different study areas for different temporal periods, we calculate different sample sizes for different spatio-temporal granularities of different modes. Primary analysis of intra-modal variability (variation within each mode of transport) suggests that motorised modes require by far much higher sample sizes than non-motorised modes, and hence, we focus on bus and car modes for this study. We use road network journey time data for the city of London to calculate appropriate sample sizes for travel surveys designed either for a specific period-of-the-day, day-of-the-week or month-of-the-year. We also use different transportation analysis zones (central, inner and outer London) to demonstrate the spatial variability of the data over these different survey durations. Then we finally calculate and analyse the range of required sample sizes for GPS-based travel surveys that are aimed at different spatio-temporal granularities in which the study is being conducted for different motorised modes of transport.