Correcting for selection bias after conditioning on a sum score in the Ising model

Area of Science:

• Psychological network analysis
• Statistical modeling
• Network science

Background:

• Psychological studies often select samples based on sum scores (e.g., symptom severity).
• This sampling method can introduce selection bias if the sum score does not perfectly represent the population of interest.
• The Ising model is a popular network model for binary data used in psychological research.

Purpose of the Study:

• To propose a statistical correction for selection bias in Ising models.
• To enable accurate estimation of network structures when samples are selected based on sum scores.
• To provide methods for obtaining full population estimates or improved subpopulation estimates.

Main Methods:

• Developed a correction method for selection bias in the Ising model.
• Validated the correction using a simulation study with node-wise regression and multivariate estimation.
• Applied the correction to empirical data on major depression symptoms from the National Comorbidity Study Replication.

Main Results:

• The proposed correction successfully recovered the network structure of the intended population after sum score selection.
• Both node-wise regression and multivariate estimation approaches demonstrated the effectiveness of the correction.
• The method was implemented in popular R packages (IsingFit, IsingSampler, psychonetrics, bootnet) for practical application.

Conclusions:

• The developed correction effectively addresses selection bias in Ising models.
• This approach enhances the accuracy of network structure estimation in psychological research.
• The implementation in R packages facilitates the practical use of this correction for researchers.