Deployment considerations for alongitudinal smartphone travel diary study

The concurrent rise of smartphone penetration and mobile sensor capability has led to increased attention on their use in mobility data collection. There appears to be promising roles for mobile device surveys to supplement or replace traditional travel surveys. While the trajectory is clear, the technology has revealed new challenges to be addressed, before the data can be regarded as robust and stable. Issues such as willingness to participate, potential biases and quality of data acquisition require study to assess the new bounds. Traditionally qualitative data collection, such as travel purpose or mode of transportation, shows real potential for improving respondent workload, but the field requires robust studies demonstrating compatibility. In order to build the groundwork for this new era in mobility research, Statistics Netherlands, in cooperation with the Dutch Ministry of Infrastructure and Water Management, developed a smartphone application for the collection of mobility data for a field test. In November 2018, 1902 respondents received letters requesting that they install an application to track their travel behaviors. In total, 674 users installed the app, of which 576 provided location data. Alongside the collection of location data, users were also requested to provide qualitative descriptions for method of transportation and purpose of visit. A set of questions requesting daily feedback on accuracy and a follow-up interview on user experience with the app were included to assess data quality and improve future applications. The goals of the field test were varied and included analysis of willingness to participate conditional on varying incentives, identification of optimal parameters for stop definition, assessing accuracy of automated transportation mode identification, and establishing criteria for alignment between data collection modes. Within this paper, the authors identify and propose solutions for the issue of non-response bias occurring at various levels within the data. Two features of the authors sample enable the authors to both identify sources of bias and attempt to correct for them. First, the authors respondents were randomly selected from the Dutch population register. This register contains comprehensive demographic information that the authors are able to link to both non-responders as well as responders. Additionally, half of the authors sample were selected from participants in a travel-diary study occurring in the two to three months immediately preceding the application-based study. The authors are therefore able to link data on detailed travel behavior both to participants and non-respondents in order to estimate the degree to which travel behaviors differ. The authors hope to additionally propose methodology to account for this in analyses. Further, the authors found meaningfully differential response rates across varying level of incentives which can be used to guide effective incentive structuring for future projects in order to reduce overall non-response. Secondly, the authors aim to demonstrate methodology for algorithmic determination of transportation modes with the ultimate goal of in-app detection to reduce overall respondent burden. The data collected within the authors app provide the authors with a near-continuous data stream of relatively precise locations, enabling calculation of track features such as speed, trajectory, location in a public transport hub or travel along known tracks. The authors field test asks participants to identify mode of transportation for each identified track, allowing the authors to determine the accuracy with which the authors can replicate the qualitative labels. Lastly, the authors outline the difficulty inherent in identifying stops. While most stops are immediately self-evident, successful transition away from diary studies, which currently require respondents to self-identify places of note, is a complex task that may vary depending on unknown or unclear requirements. Identifying the correct parameters for determining what is and is not a stop is therefore a task larger than the sum of its component parts. The authors aim to provide for a discussion of the available options, using respondents' self-tagged erroneous stops as a guideline for improving methodology, and demonstrating the ways in which parameter selection can influence the number and type of identified stops.