In wjones127/thesis: Bike Routes Thesis (Title Case)

About This Research

• This work is from my undergraduate thesis at Reed College
• Was advised by Andrew Bray

Ride Report and the Stress Map

Three Main Challenges

(1) Develop model for data collection process (2) Adapt algorithm for map matching (3) Develop model for routes

Modeling Goals

Correct for:

• riders' differing rating criteria
• ride length
• time of day (for traffic and maybe other things)
• weather:
  • Daily: avg. temp, mean wind speed, max gust speed
  • Hourly: rainfall during ride, rainfall past 4 hours

How to model?

• weather + ride length with linear predictors
• riders with random intercepts
• time of day effect with cyclic splines (by weekday/weekend)

A Logistic Regression Model

• $y_i$ are ratings
• $X$ are ride predictors
• $j[i]$ is the rider of the $i$th ride
• $t_i$ is the start time of ride $i$ in hours
• $w_i$ is an indicator for weekend

$$y_i \sim \text{Bernoulli}(p_i),$$ $$\text{logit}(p_i) = \alpha_{j[i]} + \beta \cdot X_i + s_{w_i}(t_i)$$

• $s_{w_i}(t)$ are two cyclic cubic splines, for weekdays and weekends, w/ knots at every 3 hours.
• Model fit with gamm4 package in R

Linear predictors

Bike Lanes Flood!

Photo by Richard Drdul. (https://www.flickr.com/photos/drdul/177247505)

Random Intercepts for Riders

The random intercepts were very informative for the model:

$\alpha_{j[i]}$	$\beta X_i$	$s_{w_i} (t_i)$	AIC
X	X	X	9096
	X	X	10755
X		X	9127
X	X		9153

• Maybe: Effect of cyclists' typical route absorbed by intercept

Cyclic Splines for Time of Day

Time of Day and Riders

Conclusion

• This is a start towards a goal of building a better "Stress Map""

Future models should:

• Make use of the rider intercepts, time of day splines, and weather data
• Use more granular weather data
• Be aware of the correlation between rider intercepts, time of day, and route

wjones127/thesis documentation built on May 4, 2019, 7:34 a.m.