A new study says ride-sharing apps cut cut traffic 85 percent. We’re skeptical

We’ve developed a calloused disregard for the uncritical techno-optimism that surrounds most media stories about self-driving cars and how fleets of shared-ride vehicles will neatly solve all of our urban transportation problems.

But a new story last week re-kindled our annoyance, because it so neatly captures three distinct fallacies that suggest that fleets of shared autonomous vehicles, if actually deployed in the real world, would produce a dramatically different outcome than the one imagined.  The story in question last week described a new study which reportedly proved that ride-hailing apps could reduce traffic congestion by 85 percent. The headline at Mashable was typical:

And don’t blame the writers at Mashable. Their interpretation closely mirrors a florid press release produced by MIT, which summarized the study as follows: “One way to improve traffic is through ride-sharing – and a new MIT study suggests that using carpooling options from companies like Uber and Lyft could reduce the number of vehicles on the road 75 percent without significantly impacting travel time.”

The press stories were based on a new paper entitled “On-demand high-capacity ride-sharing via dynamic trip-vehicle assignment,” written by Javier Alonso-Mora, Samitha Samaranayake, Alex Wallar, Emilio Frazzoli and Danila Rus, a team of engineers and computer scientists from MIT and Cornell. Their key finding:

Our results show that 2,000 vehicles (15% of the taxi fleet) of capacity 10 or 3,000 of capacity 4 can serve 98% of the demand within a mean waiting time of 2.8 min and mean trip delay of 3.5 min.

Despite the press release announcing the study–entitled “Study: Carpooling apps could reduce traffic 75 percent”–the study isn’t referring to all traffic, its just referring to New York City’s existing taxicab fleet. What the study is really saying is that 3,000 ten passenger vehicles with a sophisticated real time ride-hailing and vehicle assignment system could provide just as many trips as 14,000 yellow taxis, many of  which are idle or simply cruising Manhattan looking for one- or two-person fares headed to a single destination. Some of the gain comes from higher occupancy (up to 10 passengers, rather than mostly a single passenger), and other of the gain comes from better dispatching (fewer miles driven with an empty vehicle). So far, so good. But going from this observation to the conclusion that this will solve our urban congestion problems quickly runs afoul of three major fallacies.

The fixed demand fallacy

Autonomous vehicle designers may be using LIDAR, imaging, vehicle-to-vehicle communication and prodigious computing power, but down deep they’re still engineers, and they’ve apparently given no thought whatsoever to induced demand.  Just as highway engineers have assumed that there’s a fixed demand for travel and that highways need to be sized accordingly, and ignored the effect of new capacity on in stimulating added travel, the MIT study assumed that the current level of taxi use exactly captures future travel demand. Its worth noting that the demand for taxis is limited, in large part, because New York City has long regulated the number of licensed cabs via its medallion system.

There’s no reason to believe the demand for 10- and 4 passenger vehicles would be restricted to just those who currently patronize cabs. Taxis handle about 360,000 rides in the Manhattan daily. About 2.8 million travel to or from Manhattan by public transit.  If their were suddenly a viable on-street ride sharing option–especially if it were cheaper–the system could have much more demand–which could swamp the congestion reducing benefits.

The big urban transportation challenge is not simply optimizing a pre-determined set of trips, its coping with the complex feedback loops that produce a fundamental law of road congestion. This study glosses over that inconvenient truth.

The big data fallacy

A big part of what propels the illusion of fixed demand is our second fallacy:  big data.  Thanks to GIS systems in taxis, cheap telecommunications, and abundant computing power, we now live in a world where we can easily access copious data on the origin and destination of everyone of the several million annual taxi trips in New York City.  While the data is massive, it isn’t infallible or immutable: it simply reflects that decisions that travelers made with a particular technology (taxis), a particular set of prices and a set of land uses and congestions levels and alternatives that were in place at the time. It may be richly detailed, but it’s dumb: it tells us nothing about how people would behave in a different set of circumstances, with different technology and different prices.  And as big as the dataset is that’s used here, it leaves out the overwhelming majority of travelers and trips in New York who travel by train, bus, bike and foot. As we’ve suggested at City Observatory, the presence of highly selective forms of “big data” is a classic “drunk under the streetlamp” problem that focuses our attention on a few selected forms of travel, to the detriment of others. Optimizing the travel system for a relatively small segment of the population–the one for which we have rich data–doesn’t prove that this will work in the real world.

The mathematical model fallacy

The third fallacy is the mathematical model fallacy. A mathematical model can be a useful tool for sussing out the scale of problems. But in this case, it involves abstracting from and greatly simplifying the nature of the system at work. Yet press accounts, like those at Mashable are awestruck by the author’s use of a mathematical model:

This study, released Monday, used a mathematical model to figure out exactly how vehicles could best meet demand through ride-sharing.

and

“There’s a mathematical model for the autonomous future.”

The authors have constructed a very sophisticated route-setting and ride matching algorithm (taking that big data on origins and destinations as a starting point) and figured out how many 10 person person and how many 2 person vehicles it might take to handle all the 14 million trips. This requires formidable math skills, to be sure. And let’s take nothing away from the authors’ technical prowess: they’ve figured out how to solve a very complicated and huge math problem almost in real time. But simply using math to model how this might work doesn’t prove that people would actually use such a system.

Consider another possibility. Suppose we banned private cars and taxis in Manhattan and increased the number and frequency of city buses five-fold.  A bus would go by every bus stop every two or three minutes. And bus travel times would be faster, because there’d be no private cars on the road. We could probably carry all those taxi trips with even fewer vehicles.  One could even construct a mathematical model to evaluate the efficiency of this system. If we extrapolate from the MIT paper, it seems likely that if  2000, 10-person vehicles or 3,000 4-person vehicles cars can eliminate 85  percent of the trips percent of the trips, then its probably not a stretch to suggest that maybe 1,000 40-person buses could do the same thing.

Would the dogs actually eat this dog food?

Its worth asking a simple real world question: Would all or many of the people currently traveling in taxis agree to travel in ten-person or even four-person shared vehicles? If all that people were paying taxi fares for was travel between points A and B, maybe so. But there are many reasons to think they wouldn’t. For starters, travelers in pooled vehicles  get a slower trip. On average their pooled trip is going to take three and half minutes longer according to the MIT study. And people taking taxis are often paying for much more than just getting from A to B. In addition to getting to their destination as quickly as possible, they may also want the amenities of a dedicated vehicle, such as privacy–they don’t want to share their ride with other persons (even if it costs less). As David King notes at CityLab, taxis are a premium service in the world of urban transportation. And especially, what riders may be paying for is greater certainty that they are getting reliable and high priority service. The authors report the average wait time for a shared vehicle would be 2.8 minutes and the average added travel time would be 3.5 minutes, but some riders would face longer waits and greater delays. And avoiding that uncertainty or variability is a big part of why people pay for a solo taxi. And finally, it may be about status–being driven as the only occupant in a car. Often times, its faster to travel between most points in Manhattan by subway, yet many people take a taxi or uber because of just these other considerations such as comfort, convenience, priority, privacy and status.

Here’s an analogy: If Per Se or Momofuku Ko (two of New York’s swankier restaurants) served all of their food cafeteria style, they could dispense with waiters and serve 200 percent more diners than they do today. A good MBA student could probably produce a pro forma that could calculate, to the penny, how much more profit the owners would make. But it’s likely that high-end restaurant patrons–like taxi customers–are paying for something more than just the basics, and that in the real world, this wouldn’t attract many customers.

Its unfortunately too easy to oversimplify the nature of the urban transportation problem. We tend to be beguiled by new technology and blinded by big data in ways that lead us to overlook some fundamental questions about, for example, geometry. As we look to implement new technologies, like autonomous vehicles and shared ride services, we need to remember some of the hard earned lessons about things like induced demand.