Using information to curb racial discrimination: Evidence from Ride Sharing

How to cite

Ozaltun, Bora; Knittel, Christopher; MacKenzie, Don; Namazu, Michiko; Svirsky, Dan; Zoepf, Stephen. 2024. "Using information to curb racial discrimination: Evidence from Ride Sharing." AEA RCT Registry. AEARCTR-0014658. https://doi.org/10.1257/rct.14658-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Bora Ozaltun* , Christopher Knittel , Don MacKenzie , Michiko Namazu , Dan Svirsky , Stephen Zoepf

* Primary investigator

Registry ID
AEARCTR-0014658
DOI
https://doi.org/10.1257/rct.14658-1.0
First registered
Last updated
Status
Completed
Trial dates
2018-09-01 – 2018-11-30
Countries
United States of America (Boston, Los Angeles, Seattle)
Registry record
http://www.socialscienceregistry.org/trials/14658

Abstract

We test whether more information about customers decreases racial bias. Our setting is the market for shared mobility services. Prior work by Ge et al. (2020) found that Uber drivers are two times more likely to cancel a ride if the passenger's name is one used predominantly by African Americans. In a randomized control trial, we test whether two alterations to the Uber platform app reduce racial discrimination. Within the standard Uber app, drivers see only the passenger's rating before accepting a ride. Once they accept the ride, they see the name of the passenger. In the first intervention, we increased the size of the font of the rating to draw attention to the quality of the passenger. In the second intervention, the passenger's name appears from the beginning. Using the control group observations, we confirmed that the more likely African Americans were to use a name, the more likely a driver cancels the ride. However, increasing the font size of the passenger's rating eliminates this racial bias. In contrast, we do not find much evidence that showing the name on the initial screen reduces or increases cancellation rates.

Intervention

Our experiment ran in Boston, Los Angeles, and Seattle from September 2018 to November 2018. Before describing the treatments, it is important to describe the standard screen Uber drivers see when offered a fare. Uber drivers see various information about the potential trip. First, they see a map of the passenger's location and the fastest route to the passenger. The text in the middle reports the type of trip it is (UberX in this case), the rating of the passenger (4.92), the time and distance to the passenger (3 minutes and 1.2 miles, respectively), and the surge price inflater of the trip (1.6x). Notably, the passenger's name is not listed. For this reason, a cancellation is required by the driver to act on their bias.

We worked with Uber to implement two treatments. The first treatment increased the font size and color of the passenger's rating. Everything else about the initial screen remained unchanged. The second treatment added the passenger's name to the initial screen (right panel), again keeping everything else the same. Randomization took place at the driver level. In each city, 3600 drivers were included in the study, with 1200 in the control group and 1200 in each treatment group. Unlike in Ge et al. (2020), we did not randomize passenger names.

Experimental design

Our experimental design is a Randomized Control Trial (RCT).

Design details

Randomization method
Randomization of drivers ex-ante by a computer at the office.
Randomization unit
The randomization unit is at the driver level.
Clusters
We run the study in 3 cities.
Observations
In each city, we have 3600 drivers. A total of 10,800 drivers.
Arms
We observe trip level data. We have 3 cities, 3600 drivers, and 100-1000 trips per driver.
Primary outcomes
Our primary outcomes are all related to trips. We observe each trip, passenger name, driver id, drive accepted, drive cancelled, as well as a set of covariates.

Cited by

No citation data available for this registration yet.

Keywords