Graphical Model Inference with Erosely Measured Data.

Area of Science:

• Statistics
• Computational Biology
• Neuroscience

Background:

• Gaussian graphical models are crucial for understanding complex systems.
• Real-world data often exhibits irregular measurements (erose data), leading to varying sample sizes across node pairs.
• Existing methods fail to address the challenges posed by erose measurements in graph inference.

Purpose of the Study:

• To develop the first inference method specifically designed for Gaussian graphical models with erose measurements.
• To propose a method that accurately characterizes varying uncertainty levels across graph edges due to irregular data.
• To provide a statistically valid approach for graph selection in the presence of erose data.

Main Methods:

• Introduced GI-JOE (Graph Inference when Joint Observations are Erose), an edge-wise inference method.
• Developed an associated False Discovery Rate (FDR) control procedure.
• The method's variance calculation for each edge is tailored to the sample sizes of its neighboring nodes.

Main Results:

• Statistical validity of the GI-JOE method and FDR control was proven under erose measurement conditions.
• The approach successfully accounts for differing uncertainty levels across graph edges.
• Simulations and a real neuroscience dataset demonstrated superior performance in graph selection compared to existing approaches.

Conclusions:

• GI-JOE offers a robust solution for Gaussian graphical model inference with irregularly measured data.
• The method's ability to handle varying sample sizes provides significant advantages in fields like genomics and neuroscience.
• This work fills a critical gap in statistical inference for complex, real-world datasets.