Identification and Estimation Of Causal Effects from Dependent Data.

Area of Science:

• Statistics and Machine Learning
• Causal Inference
• Network Analysis

Background:

• The independent and identically distributed (iid) assumption is standard but often violated in real-world data like social networks.
• Causal inference in non-iid contexts is complicated by unobserved confounding and data dependence.
• Existing methods for causal inference primarily focus on the iid assumption.

Purpose of the Study:

• To develop a general theory for causal inference in non-iid settings.
• To provide a complete algorithm for nonparametric identification in these complex models.
• To address challenges arising from unobserved confounding and data dependence.

Main Methods:

• Utilized segregated graphs, a generalization of latent projection mixed graphs, to model causal relationships in non-iid data.
• Developed a complete algorithm for nonparametric identification of causal parameters.
• Applied the techniques to a synthetic dataset simulating fake news sharing in social networks.

Main Results:

• Established a general theory for when causal inferences are possible in non-iid scenarios.
• Demonstrated statistical inference on identified causal parameters.
• Addressed scenarios with full interference, where all units are pathwise dependent.

Conclusions:

• Segregated graphs provide a powerful framework for causal inference with dependent data.
• The developed algorithm enables valid causal conclusions in complex, non-iid settings.
• The methodology is applicable to real-world problems like analyzing information diffusion in social networks.