Understanding the Determinants of H-1B Decisions: A Machine Learning Approach

Abstract

This study aims to uncover, through an interpretable machine learning framework, the compositional shift in new employee H-1B applications that were approved before and after the 2017 “Buy American and Hire American” Executive Order (EO), which caused an unprecedented surge in denial rates. Since the Trump administration did not explicitly specify any formal regulatory changes pertaining to the H-1B visa, it is unclear a priori which subgroup(s) of applicants were most affected or, perhaps, targeted by this EO. To better understand this, we generate Shapley Additive Explanation (SHAP) values from machine learning algorithms to measure the evolution in importance of application features over time. We learn that the main drivers of application outcomes shifted substantially after the EO: previously less-important features such as sponsoring firms’ characteristics (e.g., H-1B dependence), applicants’ occupation, their salary offer, country of origin, and education level became key determinants for H-1B approval after the EO. We also find evidence that the Trump administration may have been targeting Indian applicants solely based on nationality. Indian applicants, in particular, who worked as computer programmers at H-1B dependent firms, and did not hold a graduate degree faced a notably higher likelihood of H-1B visa rejection—despite having the same (or very similar) salary offers as other applicants.

Keywords: Machine learning, H-1B visa, immigration policy, Trump

JEL Classification: F22, F23, K31