Multilevel Models and Missing Data Models for Crowdsourced Bicycle Route Ratings

Crowdsourced ratings are an increasingly important data source, leveraging the abundance of internet connected consumer devices to boost sample sizes. In this paper, we examine a data set of crowdsourced bicycle route ratings in Portland, OR collected by the Ride Report app. We fit multilevel models that show ratings are best described by models with random intercepts by rider. We also show that the majority of variation in ride ratings across time of day is owed to patterns in who is riding and where, rather than any effect particular to that time of day, such as traffic. Finally, we develop models that can adjust for non-ignorable missing ride ratings, but caution that their use for inference is inappropriate until the data quality of unrated rides can be assured.