Graphical Models for Quasi-experimental Designs.

Area of Science:

• Epidemiology
• Biostatistics
• Econometrics

Background:

• Randomized controlled trials (RCTs) and quasi-experimental designs are crucial for causal effect inference.
• Designs like regression discontinuity (RD), instrumental variable (IV), and propensity score (PS) facilitate causal interpretation.
• Understanding the identifying assumptions of these designs is vital for valid causal claims.

Purpose of the Study:

• To compare the identifying assumptions of quasi-experimental designs using causal graphs.
• To illustrate how causal graph complexity increases from RCTs to RD, IV, and PS designs.
• To introduce novel graphical representations for RD and IV designs.

Main Methods:

• Utilized causal graphs to represent and compare identifying assumptions.
• Introduced limiting graphs for regression discontinuity (RD) designs.
• Developed conditional graphs for latent subgroups in instrumental variable (IV) designs (compliers, always-takers, never-takers).

Main Results:

• Causal graph complexity increases from RCTs to RD, IV, and PS designs.
• The assumptions required for causal inference become stronger as researcher control over treatment selection diminishes.
• Propensity scores (PS) function as colliders, mitigating confounding bias through collider bias.

Conclusions:

• Causal graphs provide a powerful framework for understanding and comparing quasi-experimental designs.
• The choice of design impacts the strength of its identifying assumptions.
• Propensity score methods leverage collider bias to address confounding effectively.