A rich literature has demonstrated facial attractiveness related discrimination (beauty bias) in a wide range of contexts, such as personal marketing, election voting and employment. Under a controlled lab setting, most of these studies have found that attractiveness is rewarded. However, these studies cannot imitate long-term, real world interactions and thus lack external validity. In this paper, we show how the long-term dynamics of bias can be used to identify the source of attractiveness bias in professional settings.
We investigate two sources of attractiveness bias, namely, preference based and belief based. Belief based bias against subjects exists because evaluators have group-level priors based on the subjects’ attractiveness. These priors are overcome as the evaluator obtains objective signals of performance. Preference based bias exists because evaluators have an inherent taste for a social, romantic or marital relationship with attractive subjects. We use one of the largest archival longitudinal data sets (19,893 MBA graduates) in this area of research to identify these two sources. We find that attractiveness bias leads to a 2.6%-3.6% gap over a 15-year career period. This gap is a result of a preference bias that creates an attractiveness gap of 0.64% per year. On the other hand, belief bias has no significant role in post-MBA professional careers. This is a significant finding because belief bias toward an individual can be minimized by the individuals’ performance information. However, preference based biases are much harder to remove.
Our setting presents two key challenges in working with unstructured data. First, for an individual, we observe only one current picture, which is taken up to 25 years after the start of the individual’s professional careers. We build a generative deep learning model to create life-like versions of a face, thus allowing us to emulate the employers’ perceptions of how the individual looked at a younger age. Second, individuals move across job profiles, companies and locations, thereby making it difficult to directly compare their career milestones. We construct a preference order for jobs (job rank) based on observed job switching and text-based job title similarities.
Conference on Information Systems and Technology, Houston, TX. Oct 2017. (Best Student Paper)
Marketing Science Conference, Philadelphia, PA. June 2018.