Data Fusion for Partial Identification of Causal Effects.

Area of Science:

• Data Science
• Causal Inference
• Statistics

Background:

• Data fusion enhances learning by integrating diverse data sources.
• Causal inference methods use observational data to estimate effects, but struggle with unobserved confounding and non-exchangeable counterfactuals.
• Existing methods fail when both assumptions are violated simultaneously.

Purpose of the Study:

• To propose a novel partial identification framework for causal inference under simultaneous assumption violations.
• To enable researchers to determine the direction and robustness of causal effects.
• To quantify the severity of assumption violations required to alter conclusions.

Main Methods:

• Developed a partial identification framework with interpretable sensitivity parameters.
• Derived causal effect bounds and employed doubly robust estimators.
• Utilized breakdown frontier analysis to assess the impact of assumption violations.

Main Results:

• The proposed framework successfully identifies causal effect bounds under violated assumptions.
• Breakdown frontier analysis demonstrated how conclusions change with increasing assumption violations.
• Applied to Project STAR, the analysis confirmed the robustness of classroom size effects on student performance.

Conclusions:

• The novel framework addresses limitations in causal inference when standard assumptions fail.
• The Project STAR study's conclusions are strengthened, showing robustness to potential unmeasured biases.
• This approach enhances confidence in causal findings derived from complex, imperfect data.