Algorithmic Bias? An Empirical Study into Apparent Gender-Based Discrimination in the Display of STEM Career Ads
Despite explicit intentions to be gender neutral, a Facebook ad for STEM careers was shown more often to men than women, potentially due to economic forces and competition among advertisers.
Cite this Article
Lambrecht, Anja, and Catherine E. Tucker. “Algorithmic Bias? An Empirical Study into Apparent Gender-Based Discrimination in the Display of STEM Career Ads.” SSRN, 2018.
Lambrecht, A., & Tucker, C. E. (2018). Algorithmic Bias? An Empirical Study into Apparent Gender-Based Discrimination in the Display of STEM Career Ads. SSRN.
Lambrecht, Anja, and Catherine E. Tucker. “Algorithmic Bias? An Empirical Study into Apparent Gender-Based Discrimination in the Display of STEM Career Ads.” SSRN, 2018.
Authors
This article was written by and , and was originally published by Social Science Research Network in March of 2018Algorithms, commonly used by digital platforms within organizations and companies, are automated mathematical formulas that allow computers to solve problems and accomplish tasks. Some adaptive algorithms are able to learn new information and incorporate this data into their automated decision-making. Increasingly, many digital platforms rely on adaptive algorithms in lieu of or in addition to human decision making, since algorithms are capable of very quickly processing large amounts of live data. Automated decisions have been used to provide credit scores, grant loans, give criminal sentencing, and generate personalized online browsing suggestions and social media content.
It is commonly supposed that algorithms have the potential to rid some human bias, but a general lack of transparency from institutions reliant on algorithmic decision making has sparked a growing concern in the potential for algorithms to also generate biases[1]. There is now some documented evidence of algorithmic discrimination; however, few studies have investigated why algorithmic outcomes may be biased in the first place. This study examines some potential explanations for gender discriminatory outcomes produced from advertising algorithms.
Specifically, this study analyzes the results from a field test of a STEM (Science, Technology, Engineering and Math) related advertisement on Facebook, since STEM is a field with a persistent gender gap. The authors examine the difference in likelihood that a male or female Facebook user saw this STEM ad, and then explored three potential explanations for the gender discrepancy. The first is that the algorithm learns discriminatory behavior from actual customer behavior (e.g., women were less likely to click on the ad, or women spent less time on social media platforms); the second is that the algorithm learned the behavior from other data sources (e.g., other forms of gender discrimination, which could vary by country); the third is not about a learned bias but rather that the economics underlying ad delivery drive the differences (e.g., a higher price premium to advertise to younger women).
[1]“Algorithmic Transparency”, Electronic Privacy Information Center, https://epic.org/algorithmic-transparency/).
Across 191 countries, women were significantly less likely to be exposed to the STEM ad on Facebook than their male counterparts.
- Men saw 20% more impressions of the ad than women overall, with younger women seeing the least.
- In particular, women aged 25-34 were 40% less likely than their male counterpart of the same age to see the STEM ad.
Researchers used additional non-experimental data sources to explore potential explanations for these observed discrepancies and concluded that this difference appears to be due to greater competition to show ads to younger women, driving up the costs of advertising to this demographic.
- Women spend more time on social media platforms. Among those who do see the ad, women are more likely to click on it than men. This dispels the first potential explanation that the discrimination is due to actual consumer behavior.
- The gender discrepancy held across all countries, regardless of differences in labor force participation rates, education levels, an index of gender equality, and GDP. This dispels the second potential explanation that the discrimination is due to other data sources from which the algorithm might learn about gender inequalities.
- On average, the advertising platform suggests that advertisers bid $0.05 more to advertise to women than men. Spillover effects in the advertising market that drive up the prices of views from women aged 18-35 may be responsible for bias ads being displayed more often to men.
- Data from a separate online retailer confirmed that after clicking on an ad, younger women tend to make more purchases than men. Thus, younger women are prized not only for their clicks but because those clicks convert into purchases.
- The authors confirmed similar gender differences in ad views with smaller field tests on Google AdWords, Instagram, and Twitter.
Rather than an algorithm itself discriminating against women, it appears that it may be the spillover effects from other highly competitive advertisers that drives up the price of female views and leads to greater male views in order to be cost efficient. It seems that the higher price for female views results from the higher likelihood of women, especially aged 25 to 34, to convert each view of an advertisement into an actual purchase. The role of economics forces that might unintentionally favor men implies that algorithmic transparency and gender neutrality will not suffice in addressing unequal gender outcomes.
This study used data from a field test of an advertisement on Facebook across 191 different countries. The ad promoted job opportunities and training in STEM fields and linked to a small website that provided information on the subject. This ad was intended to be gender neutral, and the ad-serving algorithm was instructed to show the ad to both women and men between the ages of 18 and 65. In order to place an ad on this platform, advertisers must set a maximum bid they are willing to pay for each view. The algorithm uses this maximum bid along with an internally assigned “quality score” for the likelihood of the ad being clicked on to determine what ad to show each user. For this study, the researchers set a maximum bid of $0.20 for each click and raised this bid accordingly for each market to achieve at least 5,000 total views per week in each country to a maximum of $0.60 for each click. Using data provided by Facebook, both linear regression (OLS) and logit regression were used to analyze the different likelihood of seeing the STEM ad, by gender and age group.
To test various hypotheses explaining the discrepancy, researchers analyzed data from several other sources including World Bank data for 80 to 90 countries (depending on the data point) to consider whether different gender equality measures (including labor force participation and education rates) might explain significant differences between men and women’s exposure to the STEM ad. Further, they examined ad data from an online retailer to confirm whether bids for women’s views (especially younger women) were in fact more expensive. They also performed smaller field tests on other social media platforms including Google, Twitter and Instagram in order to test whether the results generalized across various platforms.
Cite this Article
Lambrecht, Anja, and Catherine E. Tucker. “Algorithmic Bias? An Empirical Study into Apparent Gender-Based Discrimination in the Display of STEM Career Ads.” SSRN, 2018.
Lambrecht, A., & Tucker, C. E. (2018). Algorithmic Bias? An Empirical Study into Apparent Gender-Based Discrimination in the Display of STEM Career Ads. SSRN.
Lambrecht, Anja, and Catherine E. Tucker. “Algorithmic Bias? An Empirical Study into Apparent Gender-Based Discrimination in the Display of STEM Career Ads.” SSRN, 2018.
