Reliable network inference from unreliable data: A tutorial on latent network modeling using STRAND.

Area of Science:

• Social Sciences
• Network Analysis
• Statistical Modeling

Background:

• Social network analysis is crucial for understanding social ties, but standard self-report measures are prone to cognitive biases.
• These reporting inaccuracies, varying by individual and item, significantly impact network properties.
• Existing methods for network reconstruction often fail to adequately address these biases.

Purpose of the Study:

• To develop a statistical tool that jointly estimates reporting biases and underlying social network structures.
• To provide researchers with a method to obtain a more impartial representation of social ties.
• To introduce the R package STRAND for easier implementation of latent network models.

Main Methods:

• Development of a latent network model to simultaneously estimate reporting biases and network parameters.
• Conducting simulation experiments with network data subjected to various reporting biases.
• Validation using empirical food/money sharing data from a rural Colombian population.

Main Results:

• Reporting biases were found to strongly impact fundamental network properties.
• Traditional network reconstruction approaches (union/intersection of double-sampled data) were insufficient.
• The proposed latent network model effectively resolved the impacts of reporting biases.

Conclusions:

• Latent network models offer a superior approach to social network analysis by accounting for reporting inaccuracies.
• The STRAND R package provides a practical tool for researchers to apply these advanced models.
• Accurate social network analysis requires methods that explicitly address respondent biases.