Bridging the gap between design and analysis: randomization inference and sensitivity analysis for matched...

Area of Science:

• Causal Inference
• Observational Studies
• Statistical Methods

Background:

• Matching is a key design in observational studies for causal inference, enabling randomization inferences under the no unmeasured confounding assumption.
• Existing methods for randomization inference and sensitivity analysis in matched designs are limited, especially for treatment doses and non-binary outcomes.
• There's a need for robust methods to handle complex matched designs, treatment variations, and different null hypotheses in causal inference.

Purpose of the Study:

• To develop novel methods for randomization inference and sensitivity analysis in general matched observational studies with treatment doses.
• To extend existing causal inference techniques to accommodate continuous or ordinal treatments and various outcome types (binary, ordinal, continuous).
• To provide tools for testing both Fisher's sharp null and Neyman-type weak nulls in complex matched designs.

Main Methods:

• Proposed new methods for randomization inference applicable to general matched designs with treatment doses.
• Developed sensitivity analysis approaches for both Fisher's sharp null and Neyman-type weak nulls.
• Methods are designed for binary, ordinal, or continuous outcome variables and incorporated into the R package doseSens.

Main Results:

• The new methods provide valid randomization inference and sensitivity analysis for general matched designs with treatment doses.
• The proposed techniques effectively address limitations in existing causal inference approaches for complex scenarios.
• Simulations and a real data application demonstrate the utility and robustness of the developed methods.

Conclusions:

• The study successfully fills critical gaps in causal inference methodology for matched observational studies with treatment doses.
• The developed methods offer a versatile framework for robust causal inference across diverse study designs and outcome types.
• The R package doseSens provides accessible tools for implementing these advanced statistical techniques in practice.