Chunking: A procedure to improve naturalistic data analysis

Every year, traffic accidents are responsible for more than 1,000,000 fatalities worldwide. Understanding the causes of traffic accidents and increasing safety on the road are priority issues for both legislators and the automotive industry. Recently, in Europe, the US and Japan, significant public funding has been allocated for performing large-scale naturalistic driving studies to better understand accident causation and the impact of safety systems on traffic safety. The data provided by these naturalistic driving studies has never been available before in this quantity and comprehensiveness and it promises to support a wide variety of data analyses. The volume and variety of the data also pose substantial challenges that demand new data reduction and analysis techniques. This paper presents a general procedure for the analysis of naturalistic driving data called chunking that can support many of these analyses by increasing their robustness and sensitivity. Chunking divides data into equivalent, elementary chunks of data to facilitate a robust and consistent calculation of parameters. This procedure was applied, as an example, to naturalistic driving data from the SeMiFOT study in Sweden and compared with alternative procedures from past studies in order to show its advantages and rationale in a specific example. Our results show how to apply the chunking procedure and how chunking can help avoid bias from data segments with heterogeneous durations (typically obtained from SQL queries). Finally, this paper shows how chunking can increase the robustness of parameter calculation, statistical sensitivity, and create a solid basis for further data analyses.