A Bootstrap Method for Goodness of Fit and Model Selection with a Single Observed Network

Area of Science:

• Network science
• Statistical modeling
• Computational biology

Background:

• Network models analyze interaction data across various scientific domains.
• Statistical and mechanistic network models capture actor dependencies.
• Analyzing single network observations is statistically challenging due to intractable likelihoods.

Purpose of the Study:

• To develop a statistical method for goodness-of-fit and model selection using a single network observation.
• To overcome the limitations of intractable likelihoods in network analysis.
• To provide a flexible approach applicable to both statistical and mechanistic network models.

Main Methods:

• A subsampling bootstrap procedure is proposed for network analysis.
• The method generates flexible resampling distributions from a single observed network.
• It allows for nuanced, high-dimensional comparisons beyond point estimates.

Main Results:

• The subsampling bootstrap circumvents intractable likelihoods for single network data.
• The approach enables goodness-of-fit assessment and model selection.
• Worked examples include simulations and analysis of yeast protein-protein interaction data.

Conclusions:

• The proposed bootstrap procedure offers a flexible resampling distribution for network statistics.
• It is applicable to diverse network models, including statistical and mechanistic types.
• This method enhances the analysis of single, observed networks in various scientific fields.