When planners quantitatively prioritize transit investments, they can use either accessibility measures or ridership measures, or some combination of the two. Accessibility calculators count how easily people are able to reach opportunities (e.g. jobs) by transit, and ridership models attempt to forecast likely passengers. Historically, agencies have used both methods for different tasks, though recent criticism of ridership models coupled with the availability of new web-based accessibility calculators has made a comparison of each method important for planners to understand.
As part of Transportation Camp DC 2018, we challenged Conveyal to debate the strengths of accessibility calculators, and Anson Stewart came forward. I defended ridership models. During the debate, we showed off Conveyal's Analysis software as well as gave the first demonstration of CityCast, our web-based travel demand model that estimates ridership.
The Case for Accessibility Calculators
- Accessibility gives a holistic view of transit’s purpose -- mobility and proximity; providing option value and potential benefits, even if they are not immediately realized.
There is a clear methodology, with limited dependence on parameters/assumptions about future.
Accessibility calculators have quick run times -- enabling iteration of analysis (exploring multiple scenarios) rather than iteration within analysis (demand/assignment feedback).
Some agencies are beginning to use accessibility in setting policy (VDOT Smart Scale accessibility ranking).
Ridership models are expensive (time, cost, and human resources) to develop, calibrate, maintain, and explain.
Long ridership model run times inhibit iterative exploration of alternatives and make project prioritization difficult.
There may be biases inherent in the use of willingness-to-pay, existing trip rates, and value of time (Lucas et al. (2015), Nahmias-Biran and Shiftan (2016), Martens (2017)).
The Case for Ridership Models
Ridership models, with a complex methodology and long history in travel demand modeling, estimate how many riders are likely to use transit service. The aim is to estimate demand for service.
The output of a ridership model (system and route-level ridership, among others) is directly observable, meaning that the model can be tested, calibrated, and refined.
A ridership model's mode choice component already includes an elegant, multi-modal accessibility score (the log-sum) that weights opportunities, travel costs, and user preferences across all modes.
Funding agencies (FTA) will continue to use ridership estimates in project comparisons for the foreseeable future.
There has not been strong correlation observed between accessibility and real-world ridership.
Accessibility calculators do require arbitrary assumptions. For example, access must be calculated as something like "jobs within 30 minutes," but distance and travel time are continuous. Asserting a particular time threshold does not match the way people make decisions: some people dislike 15 minute commutes and others do not mind 60-minute ones.
The audience was engaged and asked many important questions. We also queued up some of our own discussion topics. I will highlight some of the best ones:
A real-world correlation between high accessibility scores and actual ridership has never been established. In a previous session on bus network redesign at Transportation Camp, a participant shared a story of a new bus service connecting two major suburban job centers and residential areas in between. The service scored well in terms of increasing accessibility to jobs, but the service attracted almost no riders. Ideally, a ridership model could have identified that this route did not serve demand. On the other hand, comparisons of opening-day ridership to forecasts have a mixed record.
Project prioritization and planning
Understanding future impacts of transportation infrastructure is obviously a complicated undertaking. Project prioritization exercises in planning require that complex analysis be distilled into a single number (or a few). Developing comparative metrics from accessibility calculators is more difficult than it may seem. Isochrones, which are popular accessibility calculator outputs, visualize metrics from just a single point. Because project prioritization usually requires a region-wide perspective, isochrones are difficult or impossible to use. Regional statistics (mean, median, sum) of accessibility calculations across zones or grid cells covering a region have limitations as well. For example, a new project that improves accessibility for the bottom 10% of the population will leave the median value unchanged, and the effect on the mean might not be large enough to measure meaningfully. The sum of all accessibilities does not suffer from this problem, but can easily reach billions of "accessible jobs" in a metropolitan region that in reality has only 1 or 2 million total jobs because a single job is accessible from hundreds of locations.
Though ridership models do result in a single number more directly, there are many additional data points that can be drawn from the model. Many of these values are based in theoretical mathematical constructs, and planners have historically had a difficult time communicating them to decision makers.
Euclidean distance (buffer) vs network distance
Ridership models tend to have a simplistic view of networks; not all links are included, and pedestrians "float" across space in a straight line from their homes to transit stops. This is obviously not how pedestrians travel in reality, and many new accessibility calculators use detailed street and trail networks to more accurately measure the distance. But this can in turn lead to unrealistic problems. For instance, many long-range transit planning projects run light rail or BRT in the median of expressways; in the future these stations will include new bridges or other infrastructure for pedestrian access, but in the present pedestrian network the stations appear to have poor access. In this type of future analysis, a euclidean distance calculator would work better for capturing accessibility than real network distance would.
How do you think these measures and tools could work together?
Conveyal Analysis excels at identifying locations where accessibility appears to be higher or lower than the region average. This information can inform zoning policies, or high-level transit network design. CityCast can place a simple but powerful ridership model in the hands of planners who want to see the system and route-level impact of those changes, particularly if the regional travel demand model is inaccessible or infeasible to them.
Anson and I think these tools complement one another well. What do you think?
Thank you to Kari Watkins of Georgia Tech for moderating our session.