Travel mode imputation using GPS and accelerometer data from a multi-day travel survey

Over the past decade, interest has grown in using Global Positioning System (GPS) data to augment or even to replace traditional travel survey or activity diaries. If the full potential of this new class of data is to be realized, processing techniques will need to be standardized and automated to some degree. This paper develops a multinomial logit (MNL) model to impute travel mode from GPS and hip-mounted accelerometer data. The MNL model is the workhorse of travel demand modeling, but it has rarely been applied to GPS data processing. A web-based recall survey provided over 900 trips for estimation and 500 plus trips for validation from a larger multi-day GPS travel survey in Portland, Oregon. Special attention is given to the imputation of bicycle travel, the identification of which has been given little attention in the North American context. We also apply two existing non-MNL mode imputation models to our Portland data and to compare and test the broader transferability of specific techniques. We find that the MNL model as specified performs well overall, generally outperforming competing model forms on the Portland GPS data. Transit network data and accelerometer data significantly improve model fit for specific modes. Accelerometer data is found in particular to aid model fit for bicycling; however, external validation results were less clear. No benefit is found to segmenting models by traveler age, although not all age groups were covered by the sample. The MNL model shows strong potential for automated GPS processing and, as a commonly used transportation modeling technique, should be relatively easy to implement elsewhere.