Wibisono and Wang's MIF pinpoints causal features in unstructured treatments
The MIF query identifies which text, image, or sequential features actually drive outcomes.
Kevin Christian Wibisono and Yixin Wang propose the maximally influential feature (MIF), a new causal query for unstructured treatments like text, images, and sequences of clinical decisions. Standard average treatment effects compare fixing a treatment to one exact value versus another, but in settings like course descriptions, almost no exact description recurs—making such effects unestimable and not useful since no one wants every course to have the same description. MIF instead identifies a binary feature of the treatment,
- MIF (maximally influential feature) is a new causal query for unstructured treatments like text, images, and sequential decisions.
- The paper provides identification conditions, estimation algorithms, and a nudging procedure to revise treatments for better outcomes.
- Demonstrated on text, image, and dynamic treatment sequence applications across 74 pages and 16 figures.
Why It Matters
MIF gives professionals a way to identify actionable causal drivers in complex, unstructured interventions rather than relying on average effects.