In regions without medically certified causes of death, verbal autopsy (VA) is a critical tool for ascertaining causes of death through caregiver interviews. In this talk, I will present methods for robust cause-of-death assignment using VA data, with a focus on the challenge of distribution shifts across populations. I will introduce a latent class model framework for jointly modeling VA data from multiple populations and a federated learning extension for scenarios where individual-level data cannot be shared. The federated approach is modular, computationally efficient, and compatible with a wide range of existing VA algorithms, enabling flexible deployment in real-world mortality surveillance systems. Model performance will be evaluated using two real-world VA datasets, and I will conclude with implications and lessons learned for predictive modeling in global health.