October 31, 2016

Do ride-sourcing drivers discriminate against passengers?

New research by the Sustainable Transportation Lab and collaborators at MIT and Stanford has found that racial and, to a lesser degree, gender discrimination continue to be problems in the transportation sector. There is a long and shameful history of taxis discriminating against black passengers, whether by refusing to stop for them, or avoiding entire neighborhoods altogether. Over the past few years, various writers have reported that transportation network companies (TNCs, aka ride-sourcing platforms typified by Uber and Lyft) help to mitigate this discrimination, making it easier for black travelers to get a ride. See, for example, here, here, and here. These stories tend to emphasize the improvement Uber or Lyft offers relative to taxis. But there are also stories like this one and this one, suggesting that while discrimination may have been reduced, it has not been eliminated.

Last week I wrote about a recent paper in which we found that estimated waiting times for an UberX ride tend to be shorter in lower-income neighborhoods of Seattle, and about the same (averaged over a whole day) in areas with more versus fewer minorities. This suggests that UberX drivers are not systematically avoiding these areas. In that post, I noted that “There are many ways that inequity or discrimination could arise in the provision of these services” and that although the quality of service was equitable across neighborhoods, “additional ongoing work can help to determine whether this translates into truly equitable access for all individuals.”

Yanbo Ge, Christopher Knittel, Stephen Zoepf, and I have been doing some of that additional work over the past year and a half, and our findings are reported in a new working paper available through the National Bureau of Economic Research. The work is currently undergoing peer review, and it paints a complex picture of discrimination on TNC platforms.

I want to clarify a couple of things up front:

• We believe that discrimination in this context probably results from the decentralized decisions of individual drivers. We have no reason to think that the TNCs themselves are discriminating, although there are countermeasures that TNCs could consider to mitigate discrimination.
• We are not claiming that all TNC drivers discriminate. Although we can’t rule it out, our data could well have resulted from a subset of bad actors using the TNC platforms.
• It is impossible to say in any individual case that a driver discriminated, because there are many valid reasons that a driver might decline or cancel a trip, or choose a certain route. But when we see many such decisions disproportionately affecting certain groups, that points to discrimination in the aggregate.

With that said, here is what we found. First, we noted that there are 4 main opportunities for discrimination in providing TNC services:

1. Drivers avoiding certain neighborhoods
2. Drivers declining trip requests from certain passengers, or canceling requests after accepting them
3. Drivers taking certain passengers on longer routes
4. Drivers leaving lower star ratings for some passengers

Ryan Hughes and I addressed the first of these in the paper I mentioned earlier, and the results were encouraging from an equity standpoint. The new paper addresses the remaining three opportunities for discrimination. We collected data by sending out teams of UW and MIT undergraduate students to request and take rides on pre-assigned routes around Seattle and Boston, respectively, logging data as they went.

Here is what we found in Seattle:

• Significantly longer average times for black travelers to get an UberX or Lyft trip request accepted in Seattle. This is consistent with at least some drivers declining or accepting-then-canceling requests from black travelers.
• Significantly longer waiting times for black passengers to be picked up by their UberX in Seattle.
• No significant differences between black and white travelers in the time to get a trip accepted or for the car to arrive when using the Flywheel taxi hailing app (although Uber and Lyft were still faster than Flywheel, regardless of traveler race).
• Confirming the long-established stereotype, taxis were significantly more likely to drive past black travelers than white travelers hailing them from the curb in downtown Seattle. As shown in the following plot, the first taxi stopped for white travelers almost 60% of the time, but for black travelers less than 20% of the time. And while white travelers never had more than 4 taxis pass them, black travelers had 6 or 7 taxis pass them 20% of the time.

Following the Seattle experiment, we set about collecting data in Boston, specifically focusing on the hypotheses that (1) Since Lyft drivers see a name and photo along with the trip request, they can discriminate when choosing whether or not to accept the trip, and (2) Since Uber drivers only see a name after accepting a request, they can discriminate by cancelling the trip. At least some Uber drivers seem to believe that this practice, known as “ACRO” (Accept-Cancel-Reason-Other) or “skipping” can allow them to decline trips without hurting their cancellation rate (which they need to maintain at a sufficiently high level). To rule out the possibility that differences were due to the individual research assistants, in Boston we assigned each traveler two identities: a stereotypically white name and a stereotypically black name. We found the following cancellation rates:

As predicted, there was not a statistically significant difference in the cancellation rates for passengers using “black” vs. “white” names on Lyft, since drivers do not need to accept then cancel (they can simply discriminate up front). However, there was a significant difference for UberX, with travelers facing double the cancellation rate when using their “black” names. Interestingly, we observed that in some cases, drivers would not officially cancel the trip, but would make no attempt to actually pick up the traveler using a “black” name, or would even drive in the opposite direction for 20 minutes or more, until the research assistant canceled the trip.

We also found that females were taken on significantly longer routes than males for the same origin and destination, which could contribute to higher fares, wasted time, and perhaps personal unpleasantness with overly “chatty” driver.

The final result is a bright spot: black and white travelers on both UberX and Lyft earned virtually identical star ratings over the course of the experiment, as shown below for Lyft in Seattle. Of course, these are the ratings provided by the drivers who actually transported the passengers; ratings might be different if provided by the sorts of drivers who were declining or canceling on black travelers.

In the paper we float several potential countermeasures that TNCs might adopt to make discrimination harder and/or reduce its impacts. Here, however, I want to note that there may be a tension between these two goals in some cases.

One option would be for TNCs to move toward more anonymity, like Uber, but potentially even moreso. For example, drivers and passengers might get a code number or password instead of one anothers’ names, so that drivers will not “ACRO” the passengers. However, this wouldn’t necessarily help solve the cases like the two presented above, where drivers made it all the way to the pickup point before refusing to provide services (although the sheer brazenness required might shame some drivers out of refusing service). Personally, I found it much more frustrating the one time I had a driver cancel on me right at the pickup, rather than 30 seconds or a minute after accepting my request. In any case, even if it were difficult or impossible for a driver to decline or cancel requests, they might respond instead by avoiding neighborhoods altogether, or leaving lower star ratings for black travelers, which could impact those passengers’ ability to get future rides.

An alternative approach is to move toward more up front transparency, like Lyft, showing the passenger’s name and photo to the driver at the time of the request. This would in all likelihood make it easier for drivers to discriminate, although the passenger would not know when they had been discriminated against, and their request could be quickly fulfilled by another willing driver. The increased waiting time for black versus white Lyft passengers was undetectable in our study, thought it could likely be identified with a large enough sample. Would facilitating some discrimination in order to reduce the material impacts of other discrimination be right? Is it tantamount to making a deal with the Devil? Or something in between?

Even without knowing the answers to those last questions, there are some commonsense changes that TNCs may wish to consider. These include increasing penalties for drivers declining or canceling trips (including de facto cancellations in which the driver simply makes no attempt at picking up the passenger), and auditing drivers’ pickup and cancellation records for indications of bias. The incentive to take some passengers on longer routes could be reduced by adopting pre-set fares for a given origin and destination, something Uber is already starting to do. Finally, we are sure there is a lot more that could be learned from data that is locked away inside the TNCs. But the companies – understandably – are reluctant to share it except when they have been compelled by regulators to do so. Sharing more data publicly or with credible third parties could help to improve our understanding of how well TNCs are doing at providing access for all travelers. In the meantime, travelers might need to resort to adopting pseudonyms in order to reduce the chances of a cancellation.