Mapping Cyclist Safety in New York City

Is there a relationship between bike path access and cyclist injuries?

New York City has seen an exponential increase in bicycle usage in the last 20 years. It reports up to 76% fewer injuries or deaths per 10 million miles biked since 2000, a measure of bike safety that controls for increased usage. Even so, the city has had an exceptionally deadly year for cyclists, leading many to feel the city needs to do more. There is existing evidence that (protected) biking infrastructure is an effective risk-reduction method. Some questions I seek to address:

• What is the breakdown of biking infrastructure in New York City?
• Where are injuries and deaths happening?
• Is there a relationship between bike infrastructure (or lack thereof) and cyclist injuries and deaths?
• Are there neighborhoods with less access to (protected) bike infrastructure?
  • If so, can they be predicted by socioeconomic status?

Let’s get a quick understanding of bikes in New York.

When do cyclist injuries and accidents happen?

What about historic trends?

What kinds of bike paths are available in the city?

Where do cyclist injuries and deaths occur?

An alternative map implemented in Deck.gl:

My first attempt to understand accidents spatially led me to aggregate them within the Neighborhood Tabulation Area (NTA) they occur in and divide by NTA population. This shows not just which neighborhoods have the most accidents, but which have the most per capita. It shows more clearly where accidents are happening than by mapping each individual accident, but it does not say whether a neighborhood is over-indexed on accidents from an infrastructure perspective. Neighborhood population may not be insightful because it may have little to do with how many people bike there, or the kind of infrastructure they have, or traffic speeds. I will explore this further in the future.

I then wanted to understand one of my primary questions: whether bike infrastructure could predict cyclist accidents. To do so, I combined the NYPD Motor Vehicle Collisions and NYC Street Centerline datasets to get a distinct count of bike accidents for every street segment as a dependent variable.

Two simple and proven models for regressing discrete count data are the poisson regression and negative-binomial regression. They are not without their flaws, but they are a good starting point.

Predicting Accident Count

	Poisson		Negative-Binomial
Predictors	Incidence Rate Ratios	CI	Incidence Rate Ratios	CI
No Bike Lane	0.65 ***	0.65 – 0.66	0.65 ***	0.64 – 0.66
Class I	2.52 ***	2.45 – 2.60	2.52 ***	2.33 – 2.73
Class II	3.33 ***	3.27 – 3.40	3.33 ***	3.14 – 3.54
Class III	2.96 ***	2.88 – 3.04	2.96 ***	2.74 – 3.20
Links	0.41 ***	0.36 – 0.46	0.41 ***	0.35 – 0.48
Class I, II	1.10	0.46 – 2.64	1.10	0.19 – 6.26
Class II, III	2.68 ***	2.33 – 3.07	2.68 ***	1.81 – 3.95
Stairs	0.15 ***	0.09 – 0.25	0.15 ***	0.08 – 0.28
Class I, III	2.83 ***	2.18 – 3.67	2.83 **	1.32 – 6.05
Class II, I	6.15 ***	4.58 – 8.27	6.15 **	1.79 – 21.15
Observations	119774		119774
R2 Nagelkerke	0.153		0.059
p<0.05   ** p<0.01   *** p<0.001

---

Note:

This is an ongoing project that continues to evolve as I learn. If you have any constructive criticism please feel free to reach out, especially with suggestions about traffic accident modeling methodology.