Abstract
Causal inference from observation data often assumes "strong ignorability,"that all confounders are observed. This assumption is standard yet untestable.However, many scientific studies involve multiple causes, different variableswhose effects are simultaneously of interest. We propose the deconfounder, analgorithm that combines unsupervised machine learning and predictive modelchecking to perform causal inference in multiple-cause settings. Thedeconfounder infers a latent variable as a substitute for unobservedconfounders and then uses that substitute to perform causal inference. Wedevelop theory for when the deconfounder leads to unbiased causal estimates,and show that it requires weaker assumptions than classical causal inference.We analyze its performance in three types of studies: semi-simulated dataaround smoking and lung cancer, semi-simulated data around genomewideassociation studies, and a real dataset about actors and movie revenue. Thedeconfounder provides a checkable approach to estimating close-to-truth causaleffects.