Yong Zhang
@yongzhangzzz.bsky.social
PhD student in psychometrics and statistics at University of Groningen. Interested in experience sampling methodology and intensive longitudinal data analysis.
Happy to share more — and more work on handling missing EMA data is on the way!
November 27, 2025 at 11:36 AM
Happy to share more — and more work on handling missing EMA data is on the way!
Key takeaways:
🤔 A deeper understanding of the time series is crucial before interpreting dynamic change.
💻 Data-driven model selection is limited in detecting the true form of nonstationarity.
📴 Missing EMA data remains a major barrier to reliable inference.
🤔 A deeper understanding of the time series is crucial before interpreting dynamic change.
💻 Data-driven model selection is limited in detecting the true form of nonstationarity.
📴 Missing EMA data remains a major barrier to reliable inference.
November 27, 2025 at 11:36 AM
Key takeaways:
🤔 A deeper understanding of the time series is crucial before interpreting dynamic change.
💻 Data-driven model selection is limited in detecting the true form of nonstationarity.
📴 Missing EMA data remains a major barrier to reliable inference.
🤔 A deeper understanding of the time series is crucial before interpreting dynamic change.
💻 Data-driven model selection is limited in detecting the true form of nonstationarity.
📴 Missing EMA data remains a major barrier to reliable inference.
We then applied all candidate models to the rich 239-day EMA dataset from Kossakowski et al. and confronted the practical challenges of empirical data — especially missingness. Missing EMA data makes model fitting unstable, complicates cross-validation, and can obscure meaningful changes.
November 27, 2025 at 11:36 AM
We then applied all candidate models to the rich 239-day EMA dataset from Kossakowski et al. and confronted the practical challenges of empirical data — especially missingness. Missing EMA data makes model fitting unstable, complicates cross-validation, and can obscure meaningful changes.
The main finding: the “correct” model is far from guaranteed to be selected. Performance depends heavily on the degree of nonstationarity, time-series length, and the selection method used.
November 27, 2025 at 11:36 AM
The main finding: the “correct” model is far from guaranteed to be selected. Performance depends heavily on the degree of nonstationarity, time-series length, and the selection method used.
In our two-part study, we first simulated a range of nonstationary time series and conducted model selection (information criteria, cross-validation, out-of-sample prediction) among all candidate models.
November 27, 2025 at 11:36 AM
In our two-part study, we first simulated a range of nonstationary time series and conducted model selection (information criteria, cross-validation, out-of-sample prediction) among all candidate models.
⏳Changes/nonstationarity in time series data
We’re fascinated by how dynamics change in time-series data — and how hard it is to model those changes when we don’t fully understand the underlying processes. Data-driven model selection approaches can help, but how well does it actually work?
We’re fascinated by how dynamics change in time-series data — and how hard it is to model those changes when we don’t fully understand the underlying processes. Data-driven model selection approaches can help, but how well does it actually work?
November 27, 2025 at 11:36 AM
⏳Changes/nonstationarity in time series data
We’re fascinated by how dynamics change in time-series data — and how hard it is to model those changes when we don’t fully understand the underlying processes. Data-driven model selection approaches can help, but how well does it actually work?
We’re fascinated by how dynamics change in time-series data — and how hard it is to model those changes when we don’t fully understand the underlying processes. Data-driven model selection approaches can help, but how well does it actually work?
🏡 Key take-home message:
Before building very large EMA-based networks, it’s often better to start smaller, build interpretable models, and keep evaluating whether the model matches the theory being tested (Hoekstra et al.: often they don’t, see osf.io/preprints/ps...).
Before building very large EMA-based networks, it’s often better to start smaller, build interpretable models, and keep evaluating whether the model matches the theory being tested (Hoekstra et al.: often they don’t, see osf.io/preprints/ps...).
OSF
osf.io
November 27, 2025 at 11:33 AM
🏡 Key take-home message:
Before building very large EMA-based networks, it’s often better to start smaller, build interpretable models, and keep evaluating whether the model matches the theory being tested (Hoekstra et al.: often they don’t, see osf.io/preprints/ps...).
Before building very large EMA-based networks, it’s often better to start smaller, build interpretable models, and keep evaluating whether the model matches the theory being tested (Hoekstra et al.: often they don’t, see osf.io/preprints/ps...).
🪜 We walk through both applications step-by-step and review existing single-case network studies to highlight typical choices in variables and timepoints.
November 27, 2025 at 11:33 AM
🪜 We walk through both applications step-by-step and review existing single-case network studies to highlight typical choices in variables and timepoints.
We take two perspectives:
- ensuring sufficient power for testing individual edges, and
- ensuring good predictive accuracy of the whole network to avoid overfitting.
- ensuring sufficient power for testing individual edges, and
- ensuring good predictive accuracy of the whole network to avoid overfitting.
November 27, 2025 at 11:33 AM
We take two perspectives:
- ensuring sufficient power for testing individual edges, and
- ensuring good predictive accuracy of the whole network to avoid overfitting.
- ensuring sufficient power for testing individual edges, and
- ensuring good predictive accuracy of the whole network to avoid overfitting.
In this paper, we show how to plan the required length of a single-case EMA study if the goal is to estimate a reliable VAR network. These same tools can also be used to retrospectively assess the quality of previously published single-case networks based on the time-series length.
November 27, 2025 at 11:33 AM
In this paper, we show how to plan the required length of a single-case EMA study if the goal is to estimate a reliable VAR network. These same tools can also be used to retrospectively assess the quality of previously published single-case networks based on the time-series length.
🆒 Networks! Networks?
Dynamic symptom networks hold great promise for understanding temporal processes in mental health, but their complexity (from VAR model) raises an important question: how can we ensure we are accurately recovering these dynamics? We argue it is a must to have enough timepoints.
Dynamic symptom networks hold great promise for understanding temporal processes in mental health, but their complexity (from VAR model) raises an important question: how can we ensure we are accurately recovering these dynamics? We argue it is a must to have enough timepoints.
November 27, 2025 at 11:33 AM
🆒 Networks! Networks?
Dynamic symptom networks hold great promise for understanding temporal processes in mental health, but their complexity (from VAR model) raises an important question: how can we ensure we are accurately recovering these dynamics? We argue it is a must to have enough timepoints.
Dynamic symptom networks hold great promise for understanding temporal processes in mental health, but their complexity (from VAR model) raises an important question: how can we ensure we are accurately recovering these dynamics? We argue it is a must to have enough timepoints.
Big thanks to my wonderful coauthors @jordanrvl.bsky.social @ginettelafit.bsky.social Anja Ernst, Josip Razum, Eva Ceulemans, @bringmannlaura.bsky.social and other people that make this paper as it is today❤️
March 4, 2025 at 2:33 PM
Big thanks to my wonderful coauthors @jordanrvl.bsky.social @ginettelafit.bsky.social Anja Ernst, Josip Razum, Eva Ceulemans, @bringmannlaura.bsky.social and other people that make this paper as it is today❤️