IV estimation without distributional assumptions.

Area of Science:

• Causal inference
• Econometrics
• Epidemiology

Background:

• Instrumental Variable (IV) estimation is a powerful tool for estimating causal effects.
• It addresses confounding but relies on unstated distributional assumptions, particularly linearity.
• Previous methods for binary variables (Balke and Pearl, 1997) did not extend to complex, non-binary data.

Purpose of the Study:

• To develop a novel bounding procedure for causal effects using Instrumental Variable (IV) estimation.
• To relax the reliance on distributional assumptions in IV analysis.
• To extend existing binary bounding methods to non-binary settings.

Main Methods:

• A new bounding procedure is proposed that only requires the core IV assumption of instrument exogeneity.
• The method is generalized beyond purely binary instrument, exposure, and outcome variables.
• A tuning parameter allows for a spectrum of analyses, from point estimates to unrestricted bounds.

Main Results:

• The proposed bounding procedure provides valid causal effect estimates without assuming linearity or other distributional properties.
• The method successfully extends binary bounds to non-binary scenarios.
• An illustrative analysis of a key epidemiological study shows its conclusions were sensitive to distributional assumptions.

Conclusions:

• The novel bounding procedure offers a robust alternative to traditional IV estimation by removing distributional assumptions.
• This approach enhances the reliability of causal effect estimation in observational studies.
• The method provides flexibility for incorporating subject-matter knowledge through a tuning parameter.