Sacha Epskamp
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Research Interests
Sacha Epskamp has a total of 203 co-authors. Here are the top co-authors by number of shared publications:
Adela-Maria Isvoranu
14 shared publications- Eiko I. Fried
Department of Clinical Psychology, Leiden University, Leiden, Netherlands
11 shared publications - Denny Borsboom
Department of Psychological Methods, University of Amsterdam
8 shared publications - René Freichel
Department of Psychology, University of Amsterdam, Amsterdam, The Netherlands
8 shared publications Ria H. A. Hoekstra
7 shared publications- Sverre Urnes Johnson
University of Oslo
7 shared publications
Latest Publications
- Open Access
- 31 October 2026
- Published by Center for Open Science in PsyArXiv
Depression-Anxiety Comorbidity Across Sex During COVID-19: An Evaluation of Partial Invariance with Invariance Partial Pruning
Depression and anxiety co-occur so frequently that comorbidity is more prevalent than the occurrence of either in isolation. The conventional latent-variable perspective offers limited insight into comorbidity because its inherent causal assumptions overlook causal relations among observable symptoms that are meaningful for explaining comorbidity. At the same time, comorbidity studies based on the network perspective have largely relied on cross-sectional data, which cannot capture the dynamic nature of mental disorders, or on two-wave panel data, which does not separate between- and within-variance and biases the estimation of dynamic processes. Therefore, we present a dynamic network analysis of depression-anxiety comorbidity with the panel graphical vector autoregressive model, using invariance partial pruning to consider possible network structural differences between biological sex. Results revealed the central role of fear in predicting depressive moods and anhedonia as a possible explanation of comorbidity. Anxiety symptoms temporally preceded depression symptoms. In the contemporaneous networks, nervousness and fear were connected to depressive mood. Furthermore, worry and rumination were more associated with depression, whereas beliefs about the uncontrollability of worry and threat monitoring were more associated with anxiety. No sex difference was discovered.
- 31 October 2026
- Published by Center for Open Science in PsyArXiv
Open AccessDepression-Anxiety Comorbidity Across Sex During COVID-19: An Evaluation of Partial Invariance with Invariance Partial Pruning
Depression and anxiety co-occur so frequently that comorbidity is more prevalent than the occurrence of either in isolation. The conventional latent-variable perspective offers limited insight into comorbidity because its inherent causal assumptions overlook causal relations among observable symptoms that are meaningful for explaining comorbidity. At the same time, comorbidity studies based on the network perspective have largely relied on cross-sectional data, which cannot capture the dynamic nature of mental disorders, or on two-wave panel data, which does not separate between- and within-variance and biases the estimation of dynamic processes. Therefore, we present a dynamic network analysis of depression-anxiety comorbidity with the panel graphical vector autoregressive model, using invariance partial pruning to consider possible network structural differences between biological sex. Results revealed the central role of fear in predicting depressive moods and anhedonia as a possible explanation of comorbidity. Anxiety symptoms temporally preceded depression symptoms. In the contemporaneous networks, nervousness and fear were connected to depressive mood. Furthermore, worry and rumination were more associated with depression, whereas beliefs about the uncontrollability of worry and threat monitoring were more associated with anxiety. No sex difference was discovered.
- Open Access
- 25 October 2026
- Published by Center for Open Science in PsyArXiv
The Quantitative Psychologist Is Dead, Long Live the Quantitative Psychologist
Much of what quantitative psychologists publish consists of developing statistical methods, implementing them in software, and evaluating them in simulation studies. I argue that artificial intelligence (AI) can already do much of this work, and that its share will only grow. The quantitative psychologist has always been a jack of all trades, combining programming, statistics, and psychological knowledge without being a master of any one of them. AI is now the better jack of all trades. When methods become cheap, value moves to what AI cannot supply and to what it newly makes possible. I describe three directions: collecting trustworthy data from real people, formalizing psychological theory now that computational modeling is within reach of any researcher, and rethinking methodological educationso that students learn to judge the analyses that AI will run for them. Thespecialist role built around producing methods for others will fade, whilequantitative thinking spreads through all of psychology.
- 25 October 2026
- Published by Center for Open Science in PsyArXiv
Open AccessThe Quantitative Psychologist Is Dead, Long Live the Quantitative Psychologist
Much of what quantitative psychologists publish consists of developing statistical methods, implementing them in software, and evaluating them in simulation studies. I argue that artificial intelligence (AI) can already do much of this work, and that its share will only grow. The quantitative psychologist has always been a jack of all trades, combining programming, statistics, and psychological knowledge without being a master of any one of them. AI is now the better jack of all trades. When methods become cheap, value moves to what AI cannot supply and to what it newly makes possible. I describe three directions: collecting trustworthy data from real people, formalizing psychological theory now that computational modeling is within reach of any researcher, and rethinking methodological educationso that students learn to judge the analyses that AI will run for them. Thespecialist role built around producing methods for others will fade, whilequantitative thinking spreads through all of psychology.
- Open Access
- 4 September 2026
- Published by Center for Open Science in PsyArXiv
Beyond Perceptual Expertise: Identity and Affect Pathways to Other-Race Recognition
Other-race recognition difficulties contribute to wrongful convictions and discourage interracial integration. These difficulties are typically attributed to limited perceptual expertise with other-race faces and a tendency to categorise others by race rather than as individuals. Using theory-driven variable selection and data-driven network modelling, we show that other-race recognition is not just perceptual; it is also shaped by identity and affect. Participants (N = 512) completed face recognition tasks, and measures of personality, interracial contact, ethnic-racial attitudes and identity, and emotion. Hostility impairs other-race recognition, and is itself heightened by anti-dominant ethnic-racial identity, whereas ethnic-racial awareness and intergroup affection improve recognition. Individuating experiences mattered more than contact quantity. By focusing on personal, rather than group-based, construals of race, this study reframes other-race recognition as shaped by intergroup affect and ethnic-racial identity, not perceptual or socio-cognitive processes alone. Interventions targeting hostility and ethnic-racial identity may offer promising avenues for improving other-race recognition.
- 4 September 2026
- Published by Center for Open Science in PsyArXiv
Open AccessBeyond Perceptual Expertise: Identity and Affect Pathways to Other-Race Recognition
Other-race recognition difficulties contribute to wrongful convictions and discourage interracial integration. These difficulties are typically attributed to limited perceptual expertise with other-race faces and a tendency to categorise others by race rather than as individuals. Using theory-driven variable selection and data-driven network modelling, we show that other-race recognition is not just perceptual; it is also shaped by identity and affect. Participants (N = 512) completed face recognition tasks, and measures of personality, interracial contact, ethnic-racial attitudes and identity, and emotion. Hostility impairs other-race recognition, and is itself heightened by anti-dominant ethnic-racial identity, whereas ethnic-racial awareness and intergroup affection improve recognition. Individuating experiences mattered more than contact quantity. By focusing on personal, rather than group-based, construals of race, this study reframes other-race recognition as shaped by intergroup affect and ethnic-racial identity, not perceptual or socio-cognitive processes alone. Interventions targeting hostility and ethnic-racial identity may offer promising avenues for improving other-race recognition.
- Open Access
- 29 August 2026
- Published by Center for Open Science in PsyArXiv
Model Checking for Vector Autoregressive Models
Time series have become pervasive in psychological science, and Vector Autoregressive (VAR) models are now among the most popular tools for studying within-person dynamics in such data. However, researchers rarely check systematically how well a VAR model fits their data. This is problematic because model misfit can lead both to incorrect interpretations of model parameters and to missed structure in the data that would be theoretically interesting. We provide a tutorial that explains the theory behind model checking, discusses the most common types of VAR model misspecification in psychological time series, and introduces diagnostics for detecting them using plots and simulations. In addition, we provide code to extract predictions and residuals from popular software packages, along with the new R-package VARcheck, which allows researchers to create extensive diagnostic plots with only a few lines of code. We then apply these tools to assess the fit of a multilevel VAR model estimated on a typical empirical dataset of emotion measurements from 179 people over three weeks. We conclude by discussing three complementary ways in which model checking can advance psychological science using time series.
- 29 August 2026
- Published by Center for Open Science in PsyArXiv
Open AccessModel Checking for Vector Autoregressive Models
Time series have become pervasive in psychological science, and Vector Autoregressive (VAR) models are now among the most popular tools for studying within-person dynamics in such data. However, researchers rarely check systematically how well a VAR model fits their data. This is problematic because model misfit can lead both to incorrect interpretations of model parameters and to missed structure in the data that would be theoretically interesting. We provide a tutorial that explains the theory behind model checking, discusses the most common types of VAR model misspecification in psychological time series, and introduces diagnostics for detecting them using plots and simulations. In addition, we provide code to extract predictions and residuals from popular software packages, along with the new R-package VARcheck, which allows researchers to create extensive diagnostic plots with only a few lines of code. We then apply these tools to assess the fit of a multilevel VAR model estimated on a typical empirical dataset of emotion measurements from 179 people over three weeks. We conclude by discussing three complementary ways in which model checking can advance psychological science using time series.
- Open Access
- 1 August 2026
- Published by Center for Open Science in PsyArXiv
From RI-CLPM to panelGVAR: Evaluating Stationarity in Network Modeling of Panel Data
The random-intercept cross-lagged panel model (RI-CLPM) and the panel graphical vector-autoregression model (panelGVAR) are two popular frameworks for multivariate panel data. These models differ along two dimensions, and we contribute new methodology on both. The first dimension is stationarity: the panelGVAR constrains all parameters to be equal over time and treats the first wave as endogenous, whereas the RI-CLPM, as commonly specified, leaves every parameter wave-specific. Our first contribution is a pipeline of nested models: a guarded decision tree that tests one stationarity constraint at a time and releases those that fail. The second dimension is the parameterization: our second contribution integrates the GGMs into the RI-CLPM, yielding the random-intercept cross-lagged panel network model (RI-CLPN), which standard SEM software cannot specify. Both are implemented in the R package psychonetrics. Two simulation studies evaluate theper-constraint verdicts under likelihood-ratio, AIC, and BIC criteria and network recovery under mean and variance non-stationarity. They show which violations the pipeline detects, and what ignoring them costs when a panelGVAR is fitted directly. A tutorial and two empirical examples illustrate the pipeline.
- 1 August 2026
- Published by Center for Open Science in PsyArXiv
Open AccessFrom RI-CLPM to panelGVAR: Evaluating Stationarity in Network Modeling of Panel Data
The random-intercept cross-lagged panel model (RI-CLPM) and the panel graphical vector-autoregression model (panelGVAR) are two popular frameworks for multivariate panel data. These models differ along two dimensions, and we contribute new methodology on both. The first dimension is stationarity: the panelGVAR constrains all parameters to be equal over time and treats the first wave as endogenous, whereas the RI-CLPM, as commonly specified, leaves every parameter wave-specific. Our first contribution is a pipeline of nested models: a guarded decision tree that tests one stationarity constraint at a time and releases those that fail. The second dimension is the parameterization: our second contribution integrates the GGMs into the RI-CLPM, yielding the random-intercept cross-lagged panel network model (RI-CLPN), which standard SEM software cannot specify. Both are implemented in the R package psychonetrics. Two simulation studies evaluate theper-constraint verdicts under likelihood-ratio, AIC, and BIC criteria and network recovery under mean and variance non-stationarity. They show which violations the pipeline detects, and what ignoring them costs when a panelGVAR is fitted directly. A tutorial and two empirical examples illustrate the pipeline.
- Open Access
- 1 August 2026
- Published by Center for Open Science in PsyArXiv
Stress, somatic symptoms, and stress reduction activities: A dynamic network approach
Conceptualizing psychological stress as a dynamic network of causally connected symptoms could provide valuable insight into its relationships with somatic health issues. This may also provide insight into the functioning and maintenance of stress, along with how certain daily activities reduce stress. Intensive longitudinal data were collected from 46 university students through a mobile app. Participants completed 5 surveys at fixed 3-hour intervals every day for 7 days. In the survey, stress symptom questions were adapted from the Perceived Stress Scale and somatic symptom questions were adapted from the Patient Health Questionnaire-15 (PHQ-15), while stress reducing activity measures were newly created for this study. Multilevel vector autoregression was then used to produce temporal and contemporaneous networks. In the temporal network, feedback loops among cognitive psychological symptoms predicted other psychological symptoms and some somatic symptoms. In the contemporaneous network, positive psychological symptoms, negative psychological symptoms, and somatic symptoms each formed a separate cluster. Correlations between some psychological and somatic symptoms were also observed. Stress reducing activities were not strongly related to other variables in either network. A potential activation pathway for stress was observed. Relationships between psychological and physical stress symptoms may occur rapidly. However, numerous limitations prevent the causal interpretability of results.
- 1 August 2026
- Published by Center for Open Science in PsyArXiv
Open AccessStress, somatic symptoms, and stress reduction activities: A dynamic network approach
Conceptualizing psychological stress as a dynamic network of causally connected symptoms could provide valuable insight into its relationships with somatic health issues. This may also provide insight into the functioning and maintenance of stress, along with how certain daily activities reduce stress. Intensive longitudinal data were collected from 46 university students through a mobile app. Participants completed 5 surveys at fixed 3-hour intervals every day for 7 days. In the survey, stress symptom questions were adapted from the Perceived Stress Scale and somatic symptom questions were adapted from the Patient Health Questionnaire-15 (PHQ-15), while stress reducing activity measures were newly created for this study. Multilevel vector autoregression was then used to produce temporal and contemporaneous networks. In the temporal network, feedback loops among cognitive psychological symptoms predicted other psychological symptoms and some somatic symptoms. In the contemporaneous network, positive psychological symptoms, negative psychological symptoms, and somatic symptoms each formed a separate cluster. Correlations between some psychological and somatic symptoms were also observed. Stress reducing activities were not strongly related to other variables in either network. A potential activation pathway for stress was observed. Relationships between psychological and physical stress symptoms may occur rapidly. However, numerous limitations prevent the causal interpretability of results.
- 1 August 2026
- Published by American Psychological Association (APA) in Psychological Methods
- Vol. 31 (4) , 749-767
The invariance partial pruning approach to the network comparison in time-series and panel data.
Network models in time-series and panel data are powerful tools to investigate the dynamical relations among variables. Empirical research often seeks to compare network structures across groups/individuals to understand how element-wise associations respond differently to treatments, providing a framework to explain individual heterogeneity in treatment response and the relative efficacy of different intervention approaches. However, existing methods for comparing n = 1 idiographic networks are restricted to global tests, which cannot identify the precise location of edge heterogeneity. Furthermore, there is a lack of easily applicable methods to compare networks from panel data where just a few time points are available per person. We therefore present the invariance partial pruning (IVPP) approach, which first evaluates heterogeneity globally with the network invariance test and then determines the exact locus of heterogeneity at the edge level with partial pruning. Through simulations, we discovered that the network invariance test based on Akaike Information Criterion and Bayesian Information Criterion performed well. However, at small sample sizes, Akaike Information Criterion showed inflated false positive rates and Bayesian Information Criterion showed insufficient power to detect smaller true differences. The likelihood ratio test was prone to false discovery. Comparison with the fully constrained model revealed superior performance to the fully unconstrained model. Partial pruning successfully uncovered specific edge differences with desirable sensitivity and specificity. We conclude that IVPP is an essential supplemental to existing network methodology, enabling the comparison of networks from both time-series and panel data and testing specific edge differences. We implement the algorithm in the R package IVPP. (PsycInfo Database Record (c) 2026 APA, all rights reserved).
- 1 August 2026
- Published by American Psychological Association (APA) in Psychological Methods
- Vol. 31 (4) , 749-767
The invariance partial pruning approach to the network comparison in time-series and panel data.
Network models in time-series and panel data are powerful tools to investigate the dynamical relations among variables. Empirical research often seeks to compare network structures across groups/individuals to understand how element-wise associations respond differently to treatments, providing a framework to explain individual heterogeneity in treatment response and the relative efficacy of different intervention approaches. However, existing methods for comparing n = 1 idiographic networks are restricted to global tests, which cannot identify the precise location of edge heterogeneity. Furthermore, there is a lack of easily applicable methods to compare networks from panel data where just a few time points are available per person. We therefore present the invariance partial pruning (IVPP) approach, which first evaluates heterogeneity globally with the network invariance test and then determines the exact locus of heterogeneity at the edge level with partial pruning. Through simulations, we discovered that the network invariance test based on Akaike Information Criterion and Bayesian Information Criterion performed well. However, at small sample sizes, Akaike Information Criterion showed inflated false positive rates and Bayesian Information Criterion showed insufficient power to detect smaller true differences. The likelihood ratio test was prone to false discovery. Comparison with the fully constrained model revealed superior performance to the fully unconstrained model. Partial pruning successfully uncovered specific edge differences with desirable sensitivity and specificity. We conclude that IVPP is an essential supplemental to existing network methodology, enabling the comparison of networks from both time-series and panel data and testing specific edge differences. We implement the algorithm in the R package IVPP. (PsycInfo Database Record (c) 2026 APA, all rights reserved).
- Open Access
- 31 July 2026
- Published by Center for Open Science in PsyArXiv
Dynamical Fit Index Cutoffs for Gaussian Graphical Models
As exploratory network research continues to grow, there is heightened demand for confirmatory network analysis that can validate replicability and examine hypotheses. The quality of confirmatory testing depends critically on reliable fit index cutoffs that control type I and type II errors during hypothesis testing. However, conventional cutoffs developed by Hu and Bentler (1999; HB cutoffs) are inflexible and are vulnerable to sample and model characteristics, creating a demand for more flexible dynamic fit index (DFI) cutoffs that can be adapted to researchers’ specific designs. In this study, we attempted to develop a DFI algorithm for Gaussian graphical models, a widely implemented network model, and compared the performance of DFI cutoffs with HB cutoffsin terms of their specificity and sensitivity to misspecifications. In simulation studies based on both synthetic and empirical networks, DFI provided more flexible cutoffs with a transparent, fixed power/specificity to detect correctly specified models and was often more sensitive to misspecifications, especially in empirical networks. The DFI for network models provides a more reliable and flexible criterion for confirmatory network analysis, thereby marking an essential advancement in theory and hypothesis testing in network research. The algorithm has been implemented in an open-source, user-friendly R package, netDFI.
- 31 July 2026
- Published by Center for Open Science in PsyArXiv
Open AccessDynamical Fit Index Cutoffs for Gaussian Graphical Models
As exploratory network research continues to grow, there is heightened demand for confirmatory network analysis that can validate replicability and examine hypotheses. The quality of confirmatory testing depends critically on reliable fit index cutoffs that control type I and type II errors during hypothesis testing. However, conventional cutoffs developed by Hu and Bentler (1999; HB cutoffs) are inflexible and are vulnerable to sample and model characteristics, creating a demand for more flexible dynamic fit index (DFI) cutoffs that can be adapted to researchers’ specific designs. In this study, we attempted to develop a DFI algorithm for Gaussian graphical models, a widely implemented network model, and compared the performance of DFI cutoffs with HB cutoffsin terms of their specificity and sensitivity to misspecifications. In simulation studies based on both synthetic and empirical networks, DFI provided more flexible cutoffs with a transparent, fixed power/specificity to detect correctly specified models and was often more sensitive to misspecifications, especially in empirical networks. The DFI for network models provides a more reliable and flexible criterion for confirmatory network analysis, thereby marking an essential advancement in theory and hypothesis testing in network research. The algorithm has been implemented in an open-source, user-friendly R package, netDFI.
- Open Access
- 23 July 2026
- Published by Center for Open Science in PsyArXiv
Cross-Lagged Panel Models for Studying Psychopathology: A Comparative Overview of Structural Equation and Panel Network Approaches
Researchers have increasingly adopted complex methodological approaches to investigate the co-development of symptoms over longer time frames, such as months and years. Panel studies assess a large group of individuals at multiple time points over an extended period. Various analytical approaches exist for examining the co-development of variables in panel data, yet a comprehensive overview is lacking. This study provides a detailed review and application of four key analytical approaches for examining cross-lagged associations in panel data. The four approaches include two Structural Equation Models (SEM): the Cross- Lagged Panel Model (CLPM) and the Random-Intercept Cross-Lagged Panel Model (RI- CLPM), along with two network model adaptations: Cross-Lagged Panel Network analysis (CLPN) and Panel Graphical Vector Autoregression Model (GVAR). We describe each method’s distinct characteristics, advantages, and limitations. To illustrate these, we applied these models to a panel dataset of adolescents and young adults (NSPN 2400 cohort study), examining the relationships between impulsivity and symptoms of depression and anxiety. Results showed varying temporal associations, highlighting the importance of model selection based on research objectives and data characteristics. A simulation study demonstrated that models separating within- and between-person effects (panel GVAR, RI- CLPM) reproduced true within-person temporal effects more accurately. Our review highlights the value of using multiple approaches in multiverse analyses to assess the sensitivity of findings to different analytical methods. Ultimately, the choice of analytical method greatly influences how dynamic cross-lagged processes in developmental psychopathology are interpreted, affecting the development and refinement of relevant clinical theories.
- 23 July 2026
- Published by Center for Open Science in PsyArXiv
Open AccessCross-Lagged Panel Models for Studying Psychopathology: A Comparative Overview of Structural Equation and Panel Network Approaches
Researchers have increasingly adopted complex methodological approaches to investigate the co-development of symptoms over longer time frames, such as months and years. Panel studies assess a large group of individuals at multiple time points over an extended period. Various analytical approaches exist for examining the co-development of variables in panel data, yet a comprehensive overview is lacking. This study provides a detailed review and application of four key analytical approaches for examining cross-lagged associations in panel data. The four approaches include two Structural Equation Models (SEM): the Cross- Lagged Panel Model (CLPM) and the Random-Intercept Cross-Lagged Panel Model (RI- CLPM), along with two network model adaptations: Cross-Lagged Panel Network analysis (CLPN) and Panel Graphical Vector Autoregression Model (GVAR). We describe each method’s distinct characteristics, advantages, and limitations. To illustrate these, we applied these models to a panel dataset of adolescents and young adults (NSPN 2400 cohort study), examining the relationships between impulsivity and symptoms of depression and anxiety. Results showed varying temporal associations, highlighting the importance of model selection based on research objectives and data characteristics. A simulation study demonstrated that models separating within- and between-person effects (panel GVAR, RI- CLPM) reproduced true within-person temporal effects more accurately. Our review highlights the value of using multiple approaches in multiverse analyses to assess the sensitivity of findings to different analytical methods. Ultimately, the choice of analytical method greatly influences how dynamic cross-lagged processes in developmental psychopathology are interpreted, affecting the development and refinement of relevant clinical theories.
- Open Access
- 4 July 2026
- Published by Center for Open Science in PsyArXiv
Unlocking the potential of simulation studies in network psychometrics: A tutorial
Simulation studies are a powerful tool within the field of network psychometrics, providing insight into sample size determination and best practices for achieving accurate results. However, the advantages simulation studies provide have thus far been predominately harnessed by psychological researchers with a statistical or methodological background. While user-friendly methods for simulating psychological networks have been developed, there remains a lack of comprehensive guidelines for their effective utilization. In response, this tutorial paper serves as a guide on how to set up simulation studies for three widely used network modeling techniques: the Ising model, the Gaussian Graphical Model (GGM), and the Graphical Vector Autoregressive (GVAR) model. We explain the rationale behind simulation studies and present a step-by-step roadmap to initiate one's own simulation studies. This roadmap covers all essential aspects using a practical example to guide the reader. In addition, we provide an overview of open-source software packages available for simulating network models in the statistical programming environment R, and demonstrate their practical application. In doing so, we aim to enable more researchers to explore and resolve statistical questions within network psychometrics.
- 4 July 2026
- Published by Center for Open Science in PsyArXiv
Open AccessUnlocking the potential of simulation studies in network psychometrics: A tutorial
Simulation studies are a powerful tool within the field of network psychometrics, providing insight into sample size determination and best practices for achieving accurate results. However, the advantages simulation studies provide have thus far been predominately harnessed by psychological researchers with a statistical or methodological background. While user-friendly methods for simulating psychological networks have been developed, there remains a lack of comprehensive guidelines for their effective utilization. In response, this tutorial paper serves as a guide on how to set up simulation studies for three widely used network modeling techniques: the Ising model, the Gaussian Graphical Model (GGM), and the Graphical Vector Autoregressive (GVAR) model. We explain the rationale behind simulation studies and present a step-by-step roadmap to initiate one's own simulation studies. This roadmap covers all essential aspects using a practical example to guide the reader. In addition, we provide an overview of open-source software packages available for simulating network models in the statistical programming environment R, and demonstrate their practical application. In doing so, we aim to enable more researchers to explore and resolve statistical questions within network psychometrics.