Instrumental variables as bias amplifiers with general outcome and confounding

Area of Science:

• Causal inference
• Observational studies
• Statistical modeling

Background:

• Causal inference from observational data relies on unconfoundedness, often assuming more covariates improve estimation.
• However, adjusting for certain pretreatment covariates can amplify bias in the presence of unmeasured confounding.

Purpose of the Study:

• To generalize the understanding of bias amplification beyond linear models.
• To provide a theoretical framework for when adjusting for covariates increases causal effect estimation bias.

Main Methods:

• Developed a general theory for bias amplification under monotonicity assumptions.
• Analyzed additive and multiplicative treatment models conditional on instrumental variables and confounders.

Main Results:

• Demonstrated that bias amplification occurs in a wide class of non-linear models.
• Showed that monotonicity assumptions in specific models correspond to causal diagram structures.

Conclusions:

• The common practice of adjusting for all pretreatment covariates may be detrimental when unmeasured confounding is present.
• This work extends the theory of bias amplification, offering crucial insights for robust causal inference in observational studies.