Bridging preference-based instrumental variable studies and cluster-randomized encouragement experiments: Study...

Area of Science:

• Biostatistics
• Epidemiology
• Medical Research Methodology

Background:

• Instrumental variable (IV) methods are crucial for causal inference in medical research, especially when unmeasured confounding is present.
• IVs are often applied at the cluster level (e.g., hospital or physician preference), posing unique analytical challenges.
• Existing methods may not adequately handle cluster-level IVs or account for treatment effect heterogeneity.

Purpose of the Study:

• To propose a novel framework for embedding observational instrumental variable (IV) data within a cluster-randomized encouragement experiment.
• To develop new statistical methods for causal inference in clustered settings with potential treatment effect heterogeneity.
• To introduce and validate a new estimand, the average cluster effect ratio (ACER), addressing limitations of existing methods.

Main Methods:

• Nonbipartite matching is used to embed observational IV data into a cluster-randomized encouragement experiment design.
• Formalization and examination of potential outcomes and causal assumptions specific to this design.
• Extension of testing procedures for Fisher's sharp null hypothesis and pooled effect ratio (PER); development of a novel cluster-heterogeneous proportional treatment effect model and a randomization-based testing procedure for the ACER using mixed-integer optimization.

Main Results:

• The proposed method successfully integrates cluster-level IV data into a robust experimental design.
• The new average cluster effect ratio (ACER) estimand allows for treatment heterogeneity and avoids Simpson's paradox.
• An asymptotically valid randomization-based testing procedure for the ACER is developed and demonstrated.

Conclusions:

• The proposed cluster-randomized encouragement experiment design offers a powerful approach for causal inference with cluster-level instrumental variables.
• The average cluster effect ratio (ACER) provides a more nuanced and reliable measure of treatment effects in heterogeneous clustered populations.
• The methodology is applicable to real-world medical research, as demonstrated by its application to transesophageal echocardiography use in coronary artery bypass graft surgery.