|
Abstract
|
Paul Schimek, author of the "Bicycle Driving" blog, has taken issue with one of our recent papers ("Why cities with high bicycling rates are safer for all road users"; https://doi.org/10.1016/j.jth.2019.03.004) via a Letter to the Editor to the Journal of Transport & Health. Advocates of bicycle driving (also known as vehicular cycling) generally contend that mixing bicyclists with motor vehicle traffic is preferred over bike-specific infrastructure. It is our pleasure to respond to his critique. For the sake of clarity, we will do so point by point, and in most cases, respond with exactly what was already written in the original paper.
His first issue is less with our paper and more with how the media is interpreting the results. One of our findings is that: "Better safety outcomes are instead associated with a greater prevalence of bike facilities - particularly protected and separated bike facilities - at the block group level and, more strongly so, across the overall city." His main issue seems to be that the media has sometimes focused more on the protected bike lane result, even though our paper is clear that it includes both protected and separated bike facilities in the same measure. It is fair to say that what the media focuses on is beyond our control. For instance, the last time Dr. Marshall published a paper in the Journal of Transport and Health, "Community design, street networks, and public health" (https://doi.org/10.1016/j.jth.2014.06.002), he found that more compact street networks were associated with better health outcomes. Some media highlighted that research under titles such as "New Study Confirms: The Suburbs Make You Fat" (www.treehugger.com/urban-design/new-study-confirms-suburbs-make-you-fat.html) despite the fact that the paper never said that, nor did the authors. That previous JTH study was cross-sectional, and the paper made it clear that the results do not suggest causality. What the paper focused on was the fact that people living in certain types of street networks tended to have significantly better health outcomes, but the media did not always present that result in the same way.
In the current longitudinal paper, we did in fact find that protected and separated bike facilities were significantly associated with fewer road fatalities. With this result, Dr. Schimek also takes issue with the relative prevalence of protected bike infrastructure during our study period. He writes that there were only 12.5 miles of protected bike lanes combined in our cities and cites the People for Bikes inventory of protected bike lanes. Even using the same People for Bikes inventory of protected bike lanes, there were 28.7 miles of protected bike lanes in our study cities by the end of the study period. Our own data collection was more thorough, however. As we described in the paper, only Portland managed a GIS database of bike infrastructure that included the year the facility was built. As a result, we relied upon "a combination of emails/phone calls with city planners and in-depth review of old bike maps and historic satellite imagery available in Google Earth". We then categorized and time stamped "each piece of bike infrastructure in each city by type (i.e. protected/separated bike facilities, bike lanes, and shared lane markings or sharrows) and the year it was built" (Marshall and Ferenchak, 2019). "When we compared our ability to discern bike infrastructure via Google Earth imagery against old bike maps, our results matched up well. During the Google Earth work, however, we noticed that some protected/separated cycle tracks, for instance, were previously bike lanes or sharrows. This led us to perform the same satellite imagery review for Portland as well. After categorizing and time stamping each piece of bike infrastructure based on Federal Highway Administration (FHWA) definitions, we calculated the cumulative length and density of each facility type for each year" (Marshall and Ferenchak, 2019). This process resulted in 34.3 miles of total protected bike infrastructure across nine of our study cities by the end of our study period. Dr. Schimek also claims that we relied upon the Alliance for Walking and Biking Benchmarking Report for this data. This is not true. As discussed in our paper, the Benchmarking data was used to help us with the city selection process. When describing the city selection process, we said: "With respect to site selection, the fundamental intent was to select cities across a spectrum of bicycling, bicycling infrastructure, and road safety outcomes. Hence, we first acquired city-level American Community Survey (ACS) data so that we could assess mode share longitudinally. We then supplemented the ACS data with the data behind the Alliance for Biking and Walking Benchmarking Report" (Marshall and Ferenchak, 2019). It is not until later in the paper where we discuss the much more thorough process used to acquire data for the statistical models. Dr. Schimek also uses this same Alliance for Biking and Walking resource to say that there are over 1,300 miles of separated bike paths. Our investigation of separated facilities only included those that would be part of a bike network (excluding, for example, short, isolated facilities internal to parks that are disconnected from the larger network) and found 717.5 total miles (59.8 miles per city) by the end of the study period.
Dr. Schimek's second point is that ordinary bike lanes had the same impact on traffic safety as the protected and separated facilities, but that this result was omitted. This statement is untrue. As we directly state in the paper: "We also tested the variables representing the density of standard bike lanes in place of the protected/separated bike facility variables, and even though both standard bike lane variables were significant in the category models, they interestingly become non-significant in the full models. This suggests that improved road safety for all road users is tied to the prevalence of protected/separated bike facilities much more so than the prevalence of standard bike lanes." It seems that Dr. Schimek has confused the stepped statistical results - where we are looking only at one category of variables at a time - with the full statistical models. When we considered only the built environment category, ordinary bike lanes were indeed significantly associated with better road safety outcomes. However, when we also accounted for travel behavior and socio-demographic/socio-economic factors, that result no longer held. In the full model, the relative presence of ordinary bike lanes was non-significant, as is stated in the paper.
His third point is that significant p-values do no imply causal relationships. We agree and throughout our paper, we make it abundantly clear that we are reporting statistically significant associations and not causality. For instance in our conclusions section, we say: "our results suggest that improving bike infrastructure with more protected/separated bike facilities is significantly associated with fewer fatalities and better road safety outcomes" (Marshall and Ferenchak, 2019). At no point do we say that one causes the other. We even go on to discuss possible endogeneity and future research that controls for such in the study limitations.
The fourth point made by Dr. Schimek is that we did not control for population density or speed. Again, this is false. One of our main results is the significant association between higher intersection density - which is a variable highly correlated with population density - and safety outcomes. We say in the paper: "Higher intersection density at the block group level, a measure of street network compactness and typically illustrative of slower speed streets, was associated with fewer road fatalities as well as fewer fatal and severe crashes. Population density suggested similar trends (i.e. higher population density significantly associated with better road safety outcomes), but the variable was highly correlated with intersection density but with reduced model fit statistics" (Marshall and Ferenchak, 2019). We delved further into this connection between intersection density and speed earlier in the paper before discussing our actual speed variable: "Intersection density is a measure of street network compactness or density (Marshall and Garrick, 2012) and has been shown to be associated with road safety outcomes (Marshall and Garrick, 2011) as well as vehicle speed. Yokoo and Levinson (2016) used GPS data to study actual travel speeds in relation to street network variables and found long links to be conducive to higher speeds (Yokoo and Levinson, 2016). In order to gain a better sense of the impact of the built environment on vehicle speeds, we also collected data from an open source program called CitySpeed that aggregates an average driving speed in each city by mapping the distance and duration of over 1,000 routes in each city (Kleint, 2009). Based upon an origin-destination matrix determined by popular coffee shops and schools, this Python script then collects data for each origin-destination pair from the Google Maps API regarding average speed, distance, duration, number of turns, and the number of turns per mile. We collected data from the CitySpeed program for each of our cities and tested the city-level average driving speed result in the statistical analysis" (Marshall and Ferenchak, 2019). Although our speed variable did not turn out to significant in our statistical models, we did indeed account for vehicle speed, as discussed in the paper.
His fifth point claims that we did not account for the longitudinal nature of our data with the statistical models. However, we specifically used the statistical program Stata to conduct these analyses because it was one of the few we could find that could simultaneously conduct an analysis that was both multilevel and longitudinal. Thus, all of our statistical models were indeed specified as longitudinal. Because our statistical models were also multilevel, the data was grouped by city at the block group level of geography. This allowed us to control for differences between cities. In the paper, we stated: "Multilevel models help account for spatial autocorrelation and the idea that block group-level outcomes in the same cities share the characteristics of those cities, which would infringe upon the independence assumption of typical statistical models (Ewing et al., 2003)" (Marshall and Ferenchak, 2019). We agree that allowing for city-specific fixed-effects would have been another approach to doing so, but in this paper, we were not seeking city-level particulars. Our goal was, instead, to try to gain an overall understanding as to why our set of cities with high levels of bicycling were safer for all road users, despite the fact that bicycling is conventionally considered to be much more dangerous than driving.
The last point made by Dr. Schimek refers to the fact that better road safety for all road users does not necessarily mean better road safety outcomes for bicyclists. At no point in our paper, however, did we claim to be looking at bicyclist-specific safety outcomes. As stated numerous times, we focused on road safety as a public health impact, and more specifically, on the prevalence of road fatalities and severe injuries with respect to all road users. If one city is killing 15 residents per 100,000 residents on the roads and another is killing significantly less than half that number, we consider the second city safer and a marked improvement in road safety as a population-based public health outcome. It just so happens that the cities that are building more protected and separated bike facilities also tend to have better overall safety records. In fact, we conducted a quick analysis for the sake of this authors' response letter looking at overall road safety records based on the level of protected bike infrastructure. For this, we included only on-street protected bike infrastructure and excluded off-streets trails. Using the most recent People for Bikes inventory of protected bike lanes, we separated the 100 most populated cities with such infrastructure into quartiles based on protected bike lane density. We then looked at road fatalities for the most recent five years, from 2013 through 2017, annualized them and averaged the fatality rates. We weighted the average fatality weights by population in order to increase the representativeness of the safety impact on the overall population. For example, let's assume there is one city with 500,000 residents and a road fatality rate of 8 deaths per year per 100,000 population and a second city with 50,000 residents and a fatality rate of 15 deaths per year per 100,000 population. Simply averaging these two fatality rates would result in 11.5 fatalities per year. This would not be a fair way of considering the actual safety impact, so we instead weighted the fatality rates by population. In the above example, the average weighted fatality rate for these two cities would be 8.6 fatalities per year per 100,000 population...
Language: en |