Variational Autoencoded Multivariate Spatial Fay-Herriot Models
Zhenhua Wang, Paul A. Parker, Scott H. Holan
TL;DR
Variational Autoencoded Multivariate Spatial Fay-Herriot Models tackles the computational bottlenecks of multivariate spatial Fay-Herriot models in small-area estimation by learning spatial priors with variational autoencoders. The authors propose two variants, VSMS-FH and VGMS-FH, that replace expensive spatial precision operations with VAE-generated spatial effects, enabling scalable analysis across thousands of regions. They demonstrate the approach on five-year ACS estimates for all California census tracts, showing that VGMS-FH delivers greater flexibility and substantial efficiency gains, including feasibility for high-dimensional modeling where traditional methods fail. The method also enables reusing trained VAEs across tasks within the same geography, offering a practical pathway for official statistics and other spatial areal data domains.
Abstract
Small area estimation models are essential for estimating population characteristics in regions with limited sample sizes, thereby supporting policy decisions, demographic studies, and resource allocation, among other use cases. The spatial Fay-Herriot model is one such approach that incorporates spatial dependence to improve estimation by borrowing strength from neighboring regions. However, this approach often requires substantial computational resources, limiting its scalability for high-dimensional datasets, especially when considering multiple (multivariate) responses. This paper proposes two methods that integrate the multivariate spatial Fay-Herriot model with spatial random effects, learned through variational autoencoders, to efficiently leverage spatial structure. Importantly, after training the variational autoencoder to represent spatial dependence for a given set of geographies, it may be used again in future modeling efforts, without the need for retraining. Additionally, the use of the variational autoencoder to represent spatial dependence results in extreme improvements in computational efficiency, even for massive datasets. We demonstrate the effectiveness of our approach using 5-year period estimates from the American Community Survey over all census tracts in California.
