Race Discrimination in Internet Advertising: Evidence From a Field Experiment

How to cite

Sehgal, Neil; Svirsky, Dan. 2022. "Race Discrimination in Internet Advertising: Evidence From a Field Experiment." AEA RCT Registry. AEARCTR-0009168. https://doi.org/10.1257/rct.9168-1.0

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

Neil Sehgal* , Dan Svirsky

* Primary investigator

Registry ID
AEARCTR-0009168
DOI
https://doi.org/10.1257/rct.9168-1.0
First registered
Last updated
Status
Completed
Trial dates
2020-08-18 – 2021-09-04
Countries
United States of America ()
JEL codes
J71, J15
Registry record
http://www.socialscienceregistry.org/trials/9168

Abstract

We present the results of an experiment that documents racial bias on Facebook’s Advertising Platform. We find that darker skin complexions are penalized, leading to real economic consequences. For every $1,000 an advertiser spends on ads with models with light skin complexions, that advertiser would have to spend $1,159 to achieve the same level of engagement using a photo of a darker skin complexion model. Facebook’s budget optimization tool reinforces these viewer biases. When pictures of models with light and dark complexions are allocated a shared budget, Facebook funnels roughly 64% of the budget towards photos featuring lighter skin complexions. Our findings show how neutral algorithmic decisions can complement user bias, and also demonstrate that racial disparities persist in settings where statistical discrimination is a less natural explanation than simple animus.

Intervention

We will conduct a 2x2 experiment to test the role of racial bias in whether Instagram users like advertisements for wedding pictures. We test whether images of people with dark complexions receive less favorable engagement, relative to a baseline, when the advertisement zooms in. Consider two photographs that are highly similar but differ in the skin complexion of the people being pictured. The hypothesis is: if an Instagram ad zooms in on both pictures, in a way that makes the skin of the models take up more of the image, does this lead to racial disparities? Specifically, let P_L be equal to the Number of Likes for a photograph of a model with a lighter complexion divided by the Total Number of People Who Viewed that photograph, P_D be the same for a photograph of a model with a dark complexion, P_LZ be the same for the zoomed-in photograph of the model with a light complexion, and P_DZ be the same for a zoomed-in photograph of the model with a dark complexion. Is P_LZ – P_L > P_DZ – P_D?

Experimental design

For the likes experiment, users are assigned to one of four conditions:
1) A zoomed-out wedding photograph with people with light complexions
2) A zoomed-out wedding photograph with people with dark complexions
3) A zoomed-in wedding photograph with people with light complexions
4) A zoomed-in wedding photograph with people with dark complexions
There are three approaches to finding pairs of photographs that are similar to each other but have models of different skin complexion:
1) We could find pictures that look highly similar – same pose, same theme – but where the models photographed have different skin complexions
2) We could take a picture of someone with a light complexion and darken it using photo editing software
3) We could take a picture of someone with a dark complexion and lighten it using photo editing software
We use all three approaches, and we use each approach two times.

Design details

Randomization method
The Facebook advertisement algorithm randomly selects users to view our advertisements.
Randomization unit
Individuals are randomized to advertisements. Facebook's advertisement platform randomly selects users to view our ads based on targeting parameters.
Clusters
We will collect roughly 75,000 observations.
Observations
We will collect roughly 75,000 observations.
Arms
There are 6 sets of ads, each made of 2 zoomed out images and 2 zoomed in images. We aim for about 3,125 observations for each image, or 18,750 observations for each unique treatment (picture complexion*cropped).
Primary outcomes
The key dependent variables – P_LZ, P_L, P_DZ, and P_D – will be measured as the number of likes left for each photograph divided by the number of “reaches”, which Instagram defines as the number of people who saw the photograph in their Instagram feed.
Secondary outcomes
For our measure of potential racial disparity ( [P_LZ – P_L] – [P_DZ – P_D] ), we will have geographic breakdowns at the state level. We will calculate the geographic correlation at the state level of this measure of racial disparity with two other measures of racial disparities: the IAT self-survey of racial attitudes, and the measure of racial resentment in the ANES survey, and black maternal mortality per 1,000,000 minus white maternal mortality. Our hypothesis is that our measure of racial disparity will be positively correlated with these measures.

Cited by

No citation data available for this registration yet.

Keywords