Markov random fields
Definition
Markov random fields refers to a class of graphical models in which the conditional independence structure of multivariate data is represented through nodes and edges, where the absence of an edge between two variables indicates that those variables are conditionally independent after controlling for all remaining variables in the network. Specific instances include the Ising model for binary variables and the Gaussian Graphical Model for continuous variables, with more recent work extending the framework to ordinal data through a dedicated ordinal MRF formulation. Two sets of parameters govern any such model: the configuration of unweighted edges indicating direct connections, and the edge weight parameters capturing the strength and direction of associations between connected variable pairs. In network psychometrics, these models have been applied to symptom data for conditions such as major depressive disorder and generalized anxiety disorder, where the recovered network structure can reveal conditional dependencies among symptoms after accounting for the influence of other variables in the network.
Sources: Sekulovski et al. (2024)
Related Terms
- network psychometrics (1 shared article)
- Bayes factor (1 shared article)
- Bayesian model averaging (1 shared article)
- Bayesian variable selection (1 shared article)
- prior sensitivity (1 shared article)
Applications
Markov Random Fields and Bayesian Inference
Bayesian methods for analyzing Markov random field models require specifying prior distributions over two distinct quantities: the network structure, meaning the configuration of present edges, and the edge weight parameters associated with those edges. The inclusion Bayes factor, derived through Bayesian model averaging, serves as the primary tool for testing conditional independence between variable pairs, and its value is sensitive to the scale of the prior placed on partial correlations. Even small variations in that scale can substantially alter the Bayes factor's ability to distinguish between edge presence and edge absence.
Sources: Sekulovski et al. (2024)
Markov Random Fields and Conditional Independence Testing
The structure of a Markov random field directly encodes conditional independence relationships: two variables lacking a connecting edge are conditionally independent once all other variables in the network are held constant. Frequentist approaches combining Lasso estimation with EBIC variable selection have been the dominant analytic strategy, but these methods cannot distinguish between the absence of evidence for an edge and evidence of its absence. Bayesian approaches were developed specifically to address this limitation, enabling researchers to quantify uncertainty about both network structure and edge weights when testing for conditional independence.
Sources: Sekulovski et al. (2024)
