Investigating bias patterns in a long-term observational data set on urban mixed traffic

Commercial camera-based traffic sensors enable continuous automated collection of road user trajectories. Such data often suffer from missing values, misclassifications of road users, and erroneous positions. For technical and privacy reasons the information required to estimate or correct such errors is often not available. Here, I perform a numerical sensitivity analysis on bias patterns that can arise from these issues for a case study. I investigate the speeds at which cyclists and e-scooters travel on pavements (sidewalks) using twelve months of data. To simulate bias, I propose differential misclassification models for road user types that are informed by traffic sensor properties and take the position, movement direction, and speed of road users into account. I find that the speed difference between cyclists on the road and on the pavement are likely not robust to reasonable misclassification rates. Whilst differences in speeds between pavement and road may be small, a more robust finding is that the median speed of both cyclists and e-scooters on pavements is higher than that of pedestrians. My findings suggest that considering data quality is important, and I present a possible approach to account for road user type misclassifications.