Understanding Patenting Disparities via Causal Human+Machine Learning

We develop an empirical approach for analyzing multi-dimensional discrimination using multimodal data, combining human perception measures with language-embedding-based, nonlinear controls for latent quality to relax restrictive assumptions in causal machine learning. Applying it to the U.S. patent examination process, we find that, ceteris paribus, applications from female inventors are 1.8 percentage points less likely to be approved, and those from Black inventors are 3 percentage points less likely—inconsistent with legally prescribed criteria. Jointly studying multiple bias dimensions and their intersections for the first time, we uncover new biases, including an affiliation bias—individual inventors are disadvantaged by 6.6 percentage points relative to employees of large, public firms, a disparity larger than any demographic gap. Moreover, innovation quality, location, and other factors can mitigate or compound discrimination, and the disparities interact: for example, racial gaps vanish among public-firm employees, masking more severe discrimination against individuals. Existing theories such as homophily cannot fully explain the results, but a simple model of correlation neglect does.