A New Bayesian Foundation for Psychometric Network Modelling

Informazioni relative al progetto

BAYESIAN P-NETS

ID dell’accordo di sovvenzione: 101040876

DOI

10.3030/101040876

Data della firma CE 26 Aprile 2022

Data di avvio 1 Settembre 2022

Data di completamento 31 Agosto 2027

Finanziato da

• European Research Council (ERC)

Costo totale

€ 1 499 991,00

Contributo UE

€ 1 499 991,00

Coordinato da

UNIVERSITEIT VAN AMSTERDAM
Netherlands

Periodic Reporting for period 1 - BAYESIAN P-NETS (A New Bayesian Foundation for Psychometric Network Modelling)

Periodo di rendicontazione: 2022-09-01 al 2025-02-28

Sintesi del contesto e degli obiettivi generali del progetto

Network modeling is rapidly gaining ground as a promising way to understand psychological phenomena. A recent review identified nearly 400 papers on psychological networks published between 2008 and 2018, and an ongoing review in our project identified over 1400 of these papers between 2008 and 2023, suggesting that the literature is growing exponentially. Over the past decade, networks have become popular across the psychological sciences, but no psychological subfield has been as influenced by them as psychopathology. The network view of psychopathology posits that mental disorders arise from the causal interactions among their symptoms, and network modeling helps unravel this complex interplay.
While the past decade saw a rapidly expanding network modelling literature, methodological innovations struggled to keep pace. Reviews taking stock of the field invariably zoom in on the methodological challenges that network research faces. Three prominent concerns can be distinguished. One concern is the absence of a confirmatory scheme. Exploratory, data-driven network analyses are the norm since researchers cannot assess hypothesised network models. Another concern is the replicability of network results. Researchers publish uncertain, exploratory findings with confidence because dominant methods fail to convey the uncertainty in estimated models. The evidence plot above illustrates this point. The third concern is population heterogeneity. Current modelling approaches cannot express the individual differences underlying cross-sectional psychological data or assess if subgroups with different networks exist. But different networks may underlie the symptom interactions of subgroups.

The overall goal of this project is to establish a new research program in psychological network modeling that addresses current methodological challenges. To achieve this overall goal, the project will address three sub-goals. The first sub-goal is to develop a robust methodology that can reason with the uncertainty underlying network analysis. Building on the foundations of Bayesian inference, the project will develop advanced model-averaging methods that promise reproducible results even when there is uncertainty about the estimated network. The second sub-goal is to develop the proposed methodology for an exhaustive set of network models. To this end, the project must develop new models that are urgently needed (e.g. network models for ordinal variables). The third and final sub-goal is to compel researchers and practitioners to use the new methodology.

Lavoro eseguito dall’inizio del progetto fino alla fine del periodo coperto dalla relazione e principali risultati finora ottenuti

Sub-goal I: The Bayesian approach to psychological network analysis
We have developed a new Bayesian approach to psychological network analysis that accounts for the uncertainty underlying the choice of structure and associated parameters. Using this methodology, Bayesian model averaging can be used to analyze the structure and parameters of the network. Our review of Bayesian methods for testing the conditional independence of pairs of variables in the network showed that the proposed approach allows robust inference about the structure of the network.

Sub-goal II: Bayesian network models tailored to psychological data
The new Bayesian methodology makes use of prior assumptions about the structure of the network and its parameters. We have shown that standard assumptions from the Bayesian literature provide a reasonable starting point. However, we also want to tailor these prior assumptions to the psychological network literature. To gain insight into published network results, we conducted a large-scale reanalysis, collecting and reanalyzing datasets used for psychological network analysis. The reanalysis data provide insight into what we can expect in terms of network structure and what are reasonable parameter values. Based on the literature and what we learned from the reanalysis, we developed a new prior distribution for network structure that can account for the clustering of variables.

Most psychological data involve ordinal variables, yet there was no proper network model for their analysis. This project developed a new graphical model for ordinal psychological data. The advantage of the proposed graphical model is that it can be used to directly assess the conditional dependence structure of the network.

Sub-goal III: Compel researchers to use the new methodology
We have two strategies to encourage applied researchers to use the new methodology. The first is to develop open source software, and the second is to disseminate the new methodology through tutorials, review papers, and workshops.
We have made three contributions to open source software: First, we developed the software bgms, which incorporates all of the innovations described above. It allows full Bayesian inference on networks of binary and ordinal variables, and the latest version can also be used to analyze how such networks compare across two independent samples. Second, we created easybgm software to simplify the Bayesian analysis of graphical models using existing software packages (including bgms). This software facilitates the computation of Bayesian parameter and network estimates, hypothesis testing of network structures, and provides a wide range of plotting options. Third, the Bayesian approach to network analysis is implemented in the user-friendly JASP software. The JASP program comes with a graphical user interface and makes it easy for researchers to load their data and point and click on their desired analysis.
To encourage researchers to adopt the new methodology, we wrote a tutorial on the Bayesian approach to network analysis and a review of Bayesian methods for testing the conditional dependence or independence of pairs of variables in the network. The advantages of the new Bayesian approach over contemporary non-Bayesian approaches are central to these publications. Together, these two publications and our open source software provide a solid foundation for researchers to venture into the new Bayesian approach. To further support the dissemination of the methodology, we have given tutorials on the methodology at several international workshops on network analysis and Bayesian modeling.

Progressi oltre lo stato dell’arte e potenziale impatto previsto (incluso l’impatto socioeconomico e le implicazioni sociali più ampie del progetto fino ad ora)

The current non-Bayesian methodology forces applied network researchers to ignore the uncertainty underlying their results. Ignoring this uncertainty is risky because it leads to overconfidence. Our forthcoming reanalysis emphasizes this point and justifies skepticism about the robustness of published results. With the proposed Bayesian methodology and software implementations, network researchers can, for the first time, estimate the evidence supporting their results and use these estimates to determine robust, replicable results. This project thus provides researchers with the tools to reason about the uncertainty underlying network results and to report them with confidence.

In order for psychologists to change the way they analyze their networks, we need to develop the methodology for network models that combine the full range of variable types over the next few years. In addition, while a basic JASP implementation of the Bayesian approach to network analysis is available, we need to further develop the JASP module so that researchers can take full advantage of the new paradigm. Finally, we need to help applied researchers understand how to perform a Bayesian analysis of their network and how to report the results.