Multilevel Models and Missing Data Models for Crowdsourced Bicycle Route Ratings

May 2016

This is my undergraduate thesis, which I completed my senior year at Reed College. Andrew Bray advised this thesis. This project was done in collaborate with Knock Software, creators of the Ride Report app.

My aim in this project was to improve way the Ride Report stress map is created. The Ride Report app collects commuting bicyclists' routes and their ratings of those routes. From this crowdsourced data set of bicycle route ratings, they create a map of Portland, OR that shows each road segment colored according to the average ratings of the ride that goes through it. They call it the stress map. Though the averaging method gives results that are acceptable, we suspected that we could do better by building a probabilistic model for the effect of route on ride rating.

It turns out there are three majors hurdles in building this model: First, to develop a model that accounts for covariates other than route; second, to implement an algorithm for mapping GPS traces of routes into sequences of road segments with output suitable for modelling; and third, to develop a model that can estimate a parameter for each road segments' contribution to the ride rating. Because I was short on time, I had to focus on the first piece.

We identified three predictors besides route that the model should account for: time of day, weather and road conditions, and the variation in cyclists' tendency to give a negative rating. We propose using a logistic regression model with weather and road conditions as normal predictors, time of day fit with a cyclic cubic spline, and cyclist variation modelled with varying intercepts.

In the course of exploring this data, we also did several other analyses. We explored correcting for a non-ignorable bias in which rides were giving rates, suspecting that rides with negative experiences were more likely to be rated than those that were uneventful. Though we implemented the methods described in Ibrahim's 1996 paper, the results were not useful because we know many of the rides missing ratings are car or transit rides that were incorrectly classified as bike rides. As that classification improves, we suspect this method may produce more useful insight into the model.

File Structure

We have attempted to make this analysis reproducible, though we are not able to share the data we worked with. It contains confidential information of riders, such as their exact commutes between their home and workplace.

The repository is organized as follows:

analysis: contains exploratory analysis from original work on data manipulation and cleaning.
data: contains some data. For privacy and license reasons, most of the data is not availible in the repository.
model_iterations: contains the scripts that ran the actual models in this analysis. Most of the figures in the thesis come from these R markdown files.
logs: contains weekly logs reporting work done for the past week.
R: contains R functions written as part of the package to manipulate data
thesis: All of the writing for the thesis document. It is written entirely in R Markdown, created from an R markdown thesis template, which can be found at https://github.com/Reedies/reedtemplates)

Outline

Chapter 1: Data sources details the collection methods of the data sources used, including the Ride Report data as well as weather data.

Chapter 2: Methods outlines the main modeling methods used, including logistic regression, multilevel models, and additive models, to keep the analysis accessible for most readers.

Chapter 3: Modeling Rides and Riders compares six iteratively built statistical models, importantly demonstrating that the differences in average ride rating throughout the day can be account better by differences in the base rate at which riders give negative ratings rather than an universal daily pattern.

Chapter 4: Classifying Riders attempts to map out the distinction between different kinds of rider and then uses these classifications to improve models developed in the previous chapter.

Chapter 5: Modeling Missing Response reports the results of using expectation maximization to use ride observations that are missing a rating to reduce the potential bias in estimates in previous models and, at the same time, model the missing data mechanism.

Chapter 6: Unfinished Work: Modeling Routes outlines the data transformation challenges we faced and puts forward some suggestions for models that incorporate route information.

wjones127/thesis

ErrorLooks like something went wrong!

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Languages

Error
Looks like something went wrong!

Packages