In recent years, the network paradigm has emerged as a powerful approach for analyzing complex systems across various scientific disciplines. Within psychology, this paradigm has shown to be particularly valuable for understanding interdependencies among symptoms or behaviors, offering insights into their relational structures, intrinsic to psychological data. The paradigm was brought into the field, with a monumental spread, by the so-called network psychometrics class of models, based on Markov Random Fields, where nodes represent variables and edges represent pairwise conditional dependencies. This dissertation explores, compares, and extends a novel network approach for adequately modeling psychological assessment data. The first study compares two network-based methods for analyzing binary clinical assessment data: network psychometrics (specifically the Ising model) and latent space item response models (LSIRM). LSIRM models item responses (or symptoms) as a bipartite network of respondents and items in an unobserved metric space, where response probability decreases as the distance between respondent and item increases. Through simulation studies and empirical applications, the study highlights their value as tools for assessing symptom complex systems. The results highlight that the Ising model’s performance is sensitive to the choice of regularization methods, which depend on the assumptions about network complexity, which is subjective for real datasets, while LSIRM captures symptom dependencies robustly, regardless of network density. Beyond symptom-level networks, LSIRM provides unique insights at the patient level, as well as on the interaction between single patients and single symptoms, also offering a rich visualization toolkit that makes the extracted information easily accessible for practitioners. The second study introduces a novel approach within the LSIRM framework, tailored for Likert-scale data, built on the one-parameter graded response model (GRM): the Latent Space Graded Response Model (LSGRM). It bridges a gap in psychometric modeling: by embedding respondents and items in a shared latent space, LSGRM captures and further processes conditional dependencies (CD) in ordinal data that traditional main-effect-only models overlook, providing a significant source of information for researchers. Simulations validate its parameter recovery and CD detection, while empirical studies demonstrate its ability to generate interpretable interaction maps, individualized person profiles, and heatmaps of item similarities. Furthermore, a comparative application shows how LSGRM avoids the distortions of dichotomization yield by LSIRM, offering a better alternative for Likert-scale psychological assessments, capturing CD while preserving granularity in ordinal data. This dissertation not only explores the potential of LSIRM approaches for psychological assessment data but also introduces a new model within the framework. Its primary contributions are 1) advancing statistical approaches in measuring psychological constructs that uncover the information in item-by-person interactions and 2) providing an accessible and specific set of visualization tools to researchers and clinicians that support personalized diagnostics and inform tailored interventions.

Negli ultimi anni, il paradigma delle reti è emerso come un approccio potente per analizzare sistemi complessi in diverse discipline scientifiche. In psicologia, questo paradigma si è dimostrato particolarmente utile per comprendere le interdipendenze tra sintomi o comportamenti, offrendo una prospettiva unica sulle strutture relazionali, intrinseche ai dati psicologici. Questo paradigma è stato introdotto nel campo grazie alla cosiddetta classe di modelli della network psychometrics, che ha avuto un'ampia diffusione. Questa classe di modelli è basata sui Markov Random Fields, in cui i nodi rappresentano variabili e i legami rappresentano dipendenze condizionate tra coppie. Questa tesi esplora, confronta ed estende un nuovo approccio a reti per modellizzare in modo adeguato i dati di valutazione psicologica. Il primo studio confronta due metodi basati su reti per analizzare i dati binari delle valutazioni cliniche: la network psychometrics (in particolare il modello Ising) e il latent space item response model (LSIRM). LSIRM modella le risposte agli item (o sintomi) come una rete bipartita di soggetti e item in uno spazio metrico non osservato, dove la probabilità di risposta diminuisce con l’aumentare della distanza tra soggetto e item. Attraverso studi di simulazione e applicazioni empiriche, lo studio evidenzia il valore di questi metodi come strumenti per analizzare sistemi complessi di sintomi. I risultati mostrano che le prestazioni del modello Ising dipendono dalla scelta del metodo di regolarizzazione, basato su assunzioni sulla complessità della rete, che sono soggettive per dataset reali, mentre LSIRM cattura in modo robusto le dipendenze tra sintomi, indipendentemente dalla densità della rete. Inoltre, LSIRM fornisce informazioni uniche a livello del singolo paziente e sulle interazioni tra pazienti e sintomi, offrendo anche strumenti di visualizzazione ricchi e facilmente accessibili per i professionisti e studiosi applicati. Il secondo studio introduce un approccio innovativo all'interno del framework LSIRM, su misura per dati su scala Likert, basato sul modello Graded Response (GRM): il Latent Space Graded Response Model (LSGRM). Questo modello colma una lacuna nella modellistica psicometrica: incorporando rispondenti e item in uno spazio latente condiviso, LSGRM cattura e elabora le dipendenze condizionate (Conditional Dependencies, CD) nei dati ordinali, che i modelli tradizionali basati solo sugli effetti principali tendono a trascurare, fornendo una significativa fonte di informazioni per i ricercatori. Le simulazioni confermano la capacità del modello di stimare i parametri e rilevare le CD, mentre gli studi empirici dimostrano la sua capacità di generare mappe di interazione interpretabili, profili personalizzati e heatmap delle somiglianze tra items. Inoltre, un’applicazione a scopi comparativi mostra come LSGRM eviti le distorsioni causate dalla dicotomizzazione utilizzata da LSIRM, offrendo un'alternativa migliore per le valutazioni psicologiche su scala Likert, catturando le CD e preservando la granularità dei dati ordinali. Questa tesi non solo esplora il potenziale degli approcci LSIRM per i dati di valutazione psicologica, ma introduce anche un nuovo modello all'interno del framework. I principali contributi sono: 1) il progresso degli approcci statistici per misurare i costrutti psicologici, utilizzando le informazioni nelle interazioni tra item e soggetti, e 2) l’introduzione di strumenti di visualizzazione intuitivi e specifici per ricercatori e clinici, che supportano diagnosi personalizzate e interventi mirati.

(2025). A new network approach to psychological assessments with a latent space item response modeling framework. (Tesi di dottorato, , 2025).

A new network approach to psychological assessments with a latent space item response modeling framework

DE CAROLIS, LUDOVICA
2025

Abstract

In recent years, the network paradigm has emerged as a powerful approach for analyzing complex systems across various scientific disciplines. Within psychology, this paradigm has shown to be particularly valuable for understanding interdependencies among symptoms or behaviors, offering insights into their relational structures, intrinsic to psychological data. The paradigm was brought into the field, with a monumental spread, by the so-called network psychometrics class of models, based on Markov Random Fields, where nodes represent variables and edges represent pairwise conditional dependencies. This dissertation explores, compares, and extends a novel network approach for adequately modeling psychological assessment data. The first study compares two network-based methods for analyzing binary clinical assessment data: network psychometrics (specifically the Ising model) and latent space item response models (LSIRM). LSIRM models item responses (or symptoms) as a bipartite network of respondents and items in an unobserved metric space, where response probability decreases as the distance between respondent and item increases. Through simulation studies and empirical applications, the study highlights their value as tools for assessing symptom complex systems. The results highlight that the Ising model’s performance is sensitive to the choice of regularization methods, which depend on the assumptions about network complexity, which is subjective for real datasets, while LSIRM captures symptom dependencies robustly, regardless of network density. Beyond symptom-level networks, LSIRM provides unique insights at the patient level, as well as on the interaction between single patients and single symptoms, also offering a rich visualization toolkit that makes the extracted information easily accessible for practitioners. The second study introduces a novel approach within the LSIRM framework, tailored for Likert-scale data, built on the one-parameter graded response model (GRM): the Latent Space Graded Response Model (LSGRM). It bridges a gap in psychometric modeling: by embedding respondents and items in a shared latent space, LSGRM captures and further processes conditional dependencies (CD) in ordinal data that traditional main-effect-only models overlook, providing a significant source of information for researchers. Simulations validate its parameter recovery and CD detection, while empirical studies demonstrate its ability to generate interpretable interaction maps, individualized person profiles, and heatmaps of item similarities. Furthermore, a comparative application shows how LSGRM avoids the distortions of dichotomization yield by LSIRM, offering a better alternative for Likert-scale psychological assessments, capturing CD while preserving granularity in ordinal data. This dissertation not only explores the potential of LSIRM approaches for psychological assessment data but also introduces a new model within the framework. Its primary contributions are 1) advancing statistical approaches in measuring psychological constructs that uncover the information in item-by-person interactions and 2) providing an accessible and specific set of visualization tools to researchers and clinicians that support personalized diagnostics and inform tailored interventions.
MIGLIORATI, SONIA
JEON, MINJEONG
LSIRM; Approccio a reti; Spazio Latente; Valutazioni; Scale Likert
LSIRM; Network approach; Latent Space; Assessments; Likert Scales
M-PSI/03 - PSICOMETRIA
English
24-apr-2025
36
2023/2024
open
(2025). A new network approach to psychological assessments with a latent space item response modeling framework. (Tesi di dottorato, , 2025).
File in questo prodotto:
File Dimensione Formato  
phd_unimib_798356.pdf

accesso aperto

Descrizione: Tesi__Definitiva__De Carolis
Tipologia di allegato: Doctoral thesis
Dimensione 27.81 MB
Formato Adobe PDF
27.81 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/10281/550502
Citazioni
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
Social impact