Learning From Informatively Missing Outcomes

Missing outcomes are widespread in modern data systems, such as in healthcare or on AI platforms, and often do not occur randomly. Frequently, the missing data depends on the unobserved outcome itself and is therefore informative. If such endogenous misconfigurations are ignored or treated with standard imputation methods, biased predictions, erroneous causal inferences, and poor decisions are likely.

This presentation introduces three methodological approaches: novel estimation methods for panel data with non-random missing outcomes, representational learning with additional data modalities such as text, and the identification and use of shadow variables for improved imputation.

Together, these approaches form a unified framework for learning and decision-making in the presence of informatively missing outcomes.

Organized by:

Institute of AI in Management
LMU Munich