Network Psychometrics

Sacha Epskamp; Gunter K. J. Maris; Lourens J. Waldorp; Denny Borsboom

Source-linked AI summary

Network Psychometrics

Sacha Epskamp, Gunter K. J. Maris, Lourens J. Waldorp, Denny Borsboom

arXiv:1609.02818v2Sep 9, 2016stat.ME

TL;DR

Psychometrics has long debated how latent variables should be interpreted and how network models relate to established psychometric models. This chapter introduces Ising and related network models, demonstrates their equivalence to multidimensional item response theory models, and concludes that latent variables may also be mathematical abstractions defined on network cliques.

Problem
The chapter addresses debates over latent-variable interpretation and the unclear relationship between network models and established psychometric models.
Method
The chapter introduces Markov random fields and Ising models, establishes their equivalence with multidimensional item response theory, and reviews estimation methods.
Results
Network models can yield exactly equivalent observed-variable distributions to item response theory models, supporting a novel interpretation of latent variables as abstractions on network cliques.
Takeaways & Limitations
Latent variables can be interpreted not only as common causes or domains but also as mathematical abstractions defined on cliques of variables in a network.
Takeaways & Limitations
Reliable Ising-model estimates require considerable data because many thresholds and pairwise network parameters must be estimated.

Abstract

from arXiv · show

This chapter provides a general introduction of network modeling in psychometrics. The chapter starts with an introduction to the statistical model formulation of pairwise Markov random fields (PMRF), followed by an introduction of the PMRF suitable for binary data: the Ising model. The Ising model is a model used in ferromagnetism to explain phase transitions in a field of particles. Following the description of the Ising model in statistical physics, the chapter continues to show that the Ising model is closely related to models used in psychometrics. The Ising model can be shown to be equivalent to certain kinds of logistic regression models, loglinear models and multi-dimensional item response theory (MIRT) models. The equivalence between the Ising model and the MIRT model puts standard psychometrics in a new light and leads to a strikingly different interpretation of well-known latent variable models. The chapter gives an overview of methods that can be used to estimate the Ising model, and concludes with a discussion on the interpretation of latent variables given the equivalence between the Ising model and MIRT.

Introduction

The chapter introduces network psychometrics as an alternative to latent-variable views, emphasizing that multidimensional item response theory and Ising models can yield exactly the same distribution over observed variables.

Introduction: Network models conceptualize psychometric item responses as proxies for variables that directly interact with one another.For depression, symptoms are hypothesized to form mutually reinforcing networks rather than being determined solely by a common latent variable.
Introduction: Multidimensional item response theory models and Ising models yield exactly the same probability distribution over observed variables.The chapter exploits this equivalence to examine its consequences for network analysis.
Introduction: The chapter aims to make the connection between network models and latent-variable models explicit.It begins by introducing Markov Random Fields, the general class of models used in network analysis.

Notation

The chapter establishes notation for random variables, realizations, vectors, parameters, and observable quantities. It models a random vector of P binary variables and defines observation counts, indices, and notation for removing elements.

Notation: Random variables use capital letters, realizations use lowercase letters, vectors use boldface, and matrices use boldfaced capital letters.Roman letters denote observable variables and parameters, while Greek letters denote unspecified quantities in the notation system.
Notation: The modeled random vector X contains P binary variables taking values 1 or −1.The values represent alternatives such as correct versus incorrect, true versus false, or yes versus no.
Notation: A realization or state is denoted by x, N is the number of observations, and n(x) counts observations with response pattern x.The notation distinguishes Xi as the ith random variable from xi as its realization.
Notation: The subscripts i and j identify different random variables, with j ≠ i, while superscript −(·) denotes elements removed from a vector.For example, X−(i) removes Xi, and X−(i,j) removes Xi and Xj; corresponding lowercase notation denotes realizations.
Notation: Appendix B provides an overview of all notations used in the chapter.The notation includes both random vectors and their realizations after selected elements are removed.

Markov Random Fields

Pairwise Markov random fields represent observed variables as nodes and their connections as edges, encoding the system’s conditional independence structure. Latent common causes instead induce associations that require fully connected cliques among affected observed variables.

A graph consists of nodes V and edges E, with PMRFs using nodes for observed random variables and edges for their connections.
PMRFs encode conditional independence; in the example, X1 and X3 are independent given X2 = x2.
Observed associations can reflect causal paths, common causes, unobserved variables, or actual pairwise interactions represented by network edges.
When an unobserved common cause influences observed variables, the PMRF must contain a fully connected clique among those nodes.

Parameterizing Markov Random Fields

Pairwise Markov random fields are parameterized by positive node and pairwise potentials, normalized over all possible observations. Their PMRF form is preserved under conditioning but not marginalization, and nonadjacent nodes are conditionally independent.

Parameterization: A PMRF is parameterized as a product of strictly positive potential functions over nodes and distinct node pairs, with Z normalizing probabilities over the sample space.Node potentials encode preferences for individual states, while pairwise potentials encode preferences for joint states.
Parameterization: Pairwise potentials are relevant only for connected nodes; unconnected nodes have a constant pairwise potential equal to 1.This identification makes the pairwise contribution of unconnected nodes neutral across all realizations.
Marginalization and conditioning: Marginalizing a PMRF over nodes Xk and Xl produces a distribution that does not have the original PMRF form.Thus, a PMRF is not generally a PMRF under marginalization.
Marginalization and conditioning: Conditioning on the remaining nodes preserves the PMRF form, so a PMRF is a PMRF under conditioning.The conditional distribution is obtained by dividing the joint expression by the marginalized distribution.
Marginalization and conditioning: When nodes Xk and Xl are unconnected, their pairwise potential equals 1 and their conditional distribution factorizes into independent functions, rendering them independent.This establishes the Markov property stated for the network.

The Ising Model

The Ising model represents binary pairwise Markov random fields through threshold parameters and pairwise network interactions. Originating in statistical physics, it illustrates how local interactions can produce synchronized system-level behavior, an analogy extended to psychological variables.

Model parameters: Threshold parameters encode each node’s preference for state 1 or −1, while ωij represents an edge and pairwise interaction between nodes.A zero threshold indicates no preference; higher or lower τi favors state 1 or −1, respectively, and ωij = 0 indicates no edge.
Model parameters: Higher or lower ωij makes two connected nodes prefer the same or different states, respectively.The interaction parameter determines whether neighboring binary nodes tend to align or oppose one another.
Model formulation: The resulting binary pairwise Markov random field distribution is known as the Ising model.The model is obtained by implementing the node and pairwise potential functions in the PMRF formulation.
Worked example: In a three-node example, τ1 = τ2 = τ3 = −0.1 indicates a general preference for state −1, and one state has potential approximately 0.332.The example computes the potential for X1 = −1, X2 = 1, and X3 = −1 as exp(−0.1 + 0.1 + −0.1 + −0.5 + −0.5) ≈ 0.332.
Physical and psychological interpretation: In physics, interactions among neighboring particles can synchronize the system as a whole, producing phenomena such as spontaneous magnetization.The chapter uses this microscopic-to-macroscopic behavior to motivate analogous interpretations of interactions among psychological variables.

The Ising Model in Psychometrics

The Ising model is equivalent or closely related to prominent psychometric models, including logistic regression, loglinear analysis, and certain IRT models. Its conditional formulation yields a predictive network in which each node is predicted by its connected neighbors.

The Ising Model in Psychometrics: The Ising model is equivalent or closely related to logistic regression, loglinear analysis, and certain Item Response Theory models.The chapter discusses these relationships in sequence and also highlights earlier work connecting IRT and the Ising model.
The Ising Model in Psychometrics: Each Ising-model variable is predicted by all other variables, with connected neighbors representing the predictors of the node’s outcome.This conditional interpretation treats the Ising model as a joint distribution of response and predictor variables, forming a predictive network.
The Ising Model in Psychometrics: When loglinear models include at most second-order interaction terms, they are equivalent to the Ising model and can be represented in the same form.The passage states that this equivalence uses the same constraints typically applied in loglinear analysis.

The Relation Between the Ising Model and Item Response Theory

The Ising model is equivalent to a special case of MIRT when latent traits have a conditionally Gaussian posterior, but these traits arise from network cliques rather than hypothesized common causes. The equivalence links network structure to orthogonal latent dimensions and offers network and common-cause explanations of correlated observed variables.

The Relation Between the Ising Model and Item Response Theory: The Ising model is equivalent to a special case of the multivariate 2-parameter logistic model, with latent variables generated by network cliques rather than hypothesized common causes.This reframes the latent variables in the equivalent MIRT representation as consequences of network structure.
The Relation Between the Ising Model and Item Response Theory: A conditionally Gaussian latent-trait distribution yields a closed-form MIRT likelihood that is the Ising model.For typical latent-trait distributions, the MIRT likelihood lacks a closed form and becomes harder to approximate as the number of traits grows.
The Relation Between the Ising Model and Item Response Theory: MIRT discrimination vectors are determined by scaled eigenvectors of the Ising graph structure, making latent dimensions orthogonal and setting their number by the rank of the graph matrix.The equivalence uses a reparameterization linking difficulty and discrimination parameters to the graph matrix’s eigenstructure.
The Relation Between the Ising Model and Item Response Theory: The posterior mean of each latent trait equals a weighted sumscore, while the marginal latent-trait distribution is a mixture of equal-covariance multivariate Gaussians weighted by response-pattern probabilities.Thus, latent-trait distributions reflect the observed response patterns implied by the network model.
The Relation Between the Ising Model and Item Response Theory: In a three-node example, the network is equivalent to two latent traits representing general coherence among all nodes and contrast between the first and third nodes.The highest probabilities occur for response states in which all three nodes take the same value, corresponding to the general-coherence trait.
The Relation Between the Ising Model and Item Response Theory: Network and common-cause perspectives can provide different explanations of the same phenomenon: cliques of correlated observed variables.The network clique factorization generates a latent trait with a functional distribution through a mathematical construction.

Estimating the Ising Model

The Ising model can be estimated through likelihood-based, loglinear, pseudolikelihood, and regularized methods, but intractable normalization and limited psychometric samples create computational and statistical challenges. Regularization, especially LASSO with cross-validation or information criteria, supports interpretable sparse network estimation, while elastic net addresses nonsparse settings.

Likelihood-based estimation: Markov samplers estimate the Ising model by approximating or circumventing the parameter-dependent partition function Z, although these algorithms can remain computationally costly.Z need not be evaluated when sampling from the Ising distribution but is required for likelihood-based estimation.
Likelihood-based estimation: Because the Ising model is equivalent to a homogeneous association model, loglinear methods such as iterative proportional fitting can estimate its parameters.The chapter also introduces pseudolikelihood, which replaces the full likelihood with a product of univariate conditional likelihoods.
Pseudolikelihood estimation: Disjoint pseudolikelihood estimation repeatedly fits multiple logistic regressions, using each node as the response and all remaining nodes as predictors before combining estimates.This yields estimates for node-specific parameters and symmetric edge weights.
Regularization: Reliable estimation requires substantial data because zero cells and low-variance predictors can produce divergent parameter estimates or inflated standard errors.These problems motivate regularization for graphical models with limited observations, a common situation in psychometrics.
Regularization: LASSO penalizes absolute parameter values, shrinking some estimates exactly to zero and enabling sparse network estimation with cross-validation or information criteria.Cross-validation selects λ by predictive accuracy, whereas EBIC additionally penalizes model complexity; the LASSO assumes the population network is sparse.
Regularization: Elastic net mixes ℓ1 and ℓ2 penalties, retaining sparsity when α > 0 while averaging effects of highly correlated predictors more than LASSO.Regularized-parameter inference remains debated because LASSO parameter distributions are not well behaved, motivating repeated sample splitting and corrected p-values.

Example Analysis

The example analyzes two simulated datasets to illustrate network-model estimation, showing that regularized models outperform unregularized ones and that dataset A exhibits the expected low-rank structure under a common-cause model.

Simulation design: Two datasets each contained 500 measurements on 10 dichotomously scored items, with dataset A generated from a two-factor multidimensional Rasch model.In dataset A, the first and last five items were determined by separate factors whose levels had correlation 0.5.
Estimation results: Regularized models performed better than unregularized models in both datasets across elastic-net tuning parameters λ and α.The independence graph, with all network parameters set to zero, formed the performance plateau on the right side of the graphs.
Interpretation: Under a true common-cause model, eigenvalue decomposition corresponds to the number of latent variables, and dataset A showed the expected low-rank graph.The chapter contrasts this low-rank structure with the sparsity expected for a network representation.

The Interpretation of Latent Variables in Psychometric Models

The section contrasts common-cause and behavior-domain interpretations of latent variables, arguing that both face theoretical limitations. It presents network models as a middle ground whose interaction-based interpretation can be distinguished from statistically equivalent latent-variable models through experimental designs and connectivity changes.

The Interpretation of Latent Variables in Psychometric Models: Latent variables have long been central to psychometric models, but their theoretical status remains controversial across areas such as intelligence and personality.The debate traces back to Spearman’s conception of general intelligence as a common determinant of observed test-score differences.
The Interpretation of Latent Variables in Psychometric Models: The common-cause interpretation treats measured attributes as causes of test scores, but critics question it when the latent variable lacks substantive theoretical or empirical identification.Under this interpretation, observed response associations may be explained solely by the data structure rather than by independently established properties of the latent variable.
The Interpretation of Latent Variables in Psychometric Models: The behavior-domain interpretation avoids treating latent variables as common causes, yet it offers limited explanatory guidance about why item responses are statistically associated.It also faces requirements involving infinite domains and unidimensionality that can be difficult to satisfy for practical item sets.
The Interpretation of Latent Variables in Psychometric Models: Network approaches provide a theoretical middle ground by rejecting a causally effective latent variable and interpreting network connections as actual interactions rather than spurious correlations.This interpretation requires assuming that connections between nodes are not merely correlations and may make the theoretical status of the resulting latent variable more transparent.
The Interpretation of Latent Variables in Psychometric Models: Although Ising and IRT models are statistically equivalent for a given dataset, their common-cause and network interpretations have different causal implications.Distinguishing the models therefore requires adding quasi-experimental or experimental designs.
The Interpretation of Latent Variables in Psychometric Models: Network models additionally support predictions about externally forced changes in connection strengths, whereas reflective latent-variable models change observed associations only indirectly through latent or error variables.The network interpretation treats observed-variable associations as causally potent pathways whose altered strengths can change system behavior.

Conclusion

The conclusion presents the Ising–MIRT correspondence as a basis for novel interpretations of psychometric models and a bridge to physics. It identifies the Ising model as a canonical form for binary network psychometrics while outlining broader research opportunities.

Conclusion: The Ising–MIRT correspondence offers novel interpretations of longstanding psychometric models and connects psychometrics with the physics literature.The authors describe this connection as opening research possibilities and providing a fresh perspective on psychometric problems and challenges.
Conclusion: Research opportunities enabled by network models range from model estimation to philosophical analyses of psychological measurement.These opportunities may alter views of the foundations on which psychometric models should be built.
Conclusion: Network models can yield probability distributions exactly equivalent to those of IRT models, allowing latent variables to receive a novel interpretation.This extends interpretations of latent variables beyond common causes of item responses and sampled behavior domains.
Conclusion: The Ising model is proposed as a canonical form for network psychometrics because it handles binary data and is equivalent to well-known IRT models.It can be interpreted as a model for interacting components, an association model with at most pairwise interactions, or a joint distribution of response and predictor variables in logistic regression.
Conclusion: The Gaussian Random Field offers a related continuous-data model with a tractable normalizing constant, unlike the Ising model’s intractable partition function.In the GRF, the inverse covariance matrix is a precision matrix that can be standardized as a partial correlation matrix.

Appendix B · Glossary of Notation

The glossary defines notation for binary-variable networks, Ising-model parameters, latent-variable and loglinear models, likelihoods, and regularization controls. It also specifies dimensions, realizations, potentials, and indexing conventions used throughout the analysis.

Glossary of Notation: X is a P-dimensional random vector of binary variables, with entries in {−1, 1}, and x is a possible realization of X.P denotes the number of variables, while N denotes the number of observations.
Glossary of Notation: n(x) counts observations with response pattern x, while i, j, k, and l index variables in {1, 2, . . . , P}, with j ≠ i.The notation distinguishes observed response-pattern frequencies from variable subscripts.
Glossary of Notation: X−(i) and x−(i) denote the random vector and realization obtained by excluding Xi, whereas X−(i,j) excludes both Xi and Xj.The corresponding dimensions are P − 1 and P − 2, respectively.
Glossary of Notation: The probability function is Pr(. . .), and node and pairwise potential functions are φi(xi) and φi(xi, xj), mapping binary inputs to positive real values.The node potential has domain {−1, 1}, while the pairwise potential has domain {−1, 1}^2.
Glossary of Notation: τi is the Ising-model threshold parameter for Xi, defined as τi = ln φi(1), and τ is the vector containing these thresholds.τ contains τi as its ith element.
Glossary of Notation: ωij is the Ising-model network parameter between Xi and Xj, defined as ωij = ln φij(1, 1), while Ω contains these parameters and ωi denotes its ith row or column.Pen(ωi) is the penalization function applied to ωi.
Glossary of Notation: β denotes inverse temperature and the Hamiltonian function denotes the energy of state x in the Ising model; ν...(. . .) denotes log potential functions and M the number of latent factors.Θ and θ denote a continuous latent-variable vector and its realization, respectively.
Glossary of Notation: L(τ,Ω;x) is the likelihood based on Pr(X = x), while λ is the LASSO tuning parameter and α is the elastic-net tuning parameter.The glossary specifies λ > 0 and α ∈ (0, 1).