Supervised MELD for Multi-domain Mixed Membership Analyses

Abstract

When variables used in a mixed membership analysis can be classied into conceptu-ally distinct domains, interpretation of results is facilitated by using domain-specicmodels with a small number of imposed pure-type proles and yielding a set of Grade-of-Membership scores for each domain. We present Supervised Moment Estimationof Latent Dirichlet Models (supervised MELD) algorithms for mixed membershipmodels to be used and tested in this context. We challenge the methodology withdata sets collected over time to study malaria risk on the Brazilian Amazon frontier.By further ignoring spatial specicity and utilizing a binary outcome variable (atleast one person in a household infected with malaria during the past year), we ndsupervised MELD capable of extracting surprisingly nuanced risk patterns.