Estimating the total treatment effect in randomized experiments with unknown network structure.

Area of Science:

• Causal inference
• Experimental design
• Network analysis

Background:

• Randomized experiments are standard for estimating treatment effects across diverse scientific fields.
• Classical methods are biased by network effects (interference), where individual treatments impact neighbors' outcomes.
• Network structure is often unknown, posing a significant challenge for accurate estimation.

Purpose of the Study:

• To address bias in randomized experiments caused by unobserved network interference.
• To develop a method for unbiased estimation of total treatment effects in the presence of network effects.
• To provide a practical and theoretically sound approach for designing experiments with network interference.

Main Methods:

• Developed a potential outcomes model with heterogeneous additive network effects.
• Characterized limitations of existing methods without network knowledge.
• Proposed a novel randomized design and a simple estimator using historical baseline data.

Main Results:

• The proposed estimator yields unbiased, low-variance estimates of total treatment effects.
• The method does not require knowledge of the underlying network structure.
• Statistical guarantees are provided for a broad class of network interference models.

Conclusions:

• The developed method overcomes limitations of classical approaches in the presence of network interference.
• The new design and estimator offer a practical solution for real-world randomized experiments.
• This work is expected to significantly impact the design and analysis of experiments involving network effects.