Causal Effect Estimation with Learned Instrument Representations

Researchers have developed ZNet, a new AI model that can learn instrumental variable (IV) representations from observed data, bypassing a major roadblock in causal inference. Traditional IV methods require a pre-identified, valid instrument—often unavailable—to estimate causal effects without bias from hidden confounders. ZNet's architecture decomposes features into confounding and instrumental components, creating latent instruments when none exist. This "plug-and-play" module works with standard IV estimators, enabling causal discovery in general observational settings where key assumptions are untestable.