Pith. sign in

REVIEW 3 major objections 4 minor 31 references

Bayesian Dynamic Clustering Factor Models

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proposes a Bayesian dynamic clustering factor model that simultaneously reduces multivariate longitudinal data to latent factors, assigns each subject to a cluster at each time point, and estimates the Markov transition…

desk verdict A useful longitudinal extension of BCFM, but the Gibbs sampler omits the initial-state probability in the Z_i1 update, so the reported simulation results do not validate the stated model. read the letter →

arxiv 2505.21490 v1 pith:VKRCXJEQ submitted 2025-05-27 stat.ME

Tsering Dolkar , Marco A. R. Ferreira , Hwasoo Shin , Allison N. Tegge This is my paper
classification stat.ME MSC 62F1562H3062M05
keywords BayesiandynamicclusteringfactormodelshiddenMarkovmultivariatelongitudinaldataGibbssamplerempiricalBayesopioidusedisorderclustertransitions
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a model that does three things at once for multivariate longitudinal data: it reduces many observed variables to a few latent factors, it assigns each subject to a subgroup at every time point, and it estimates how subjects move among subgroups over time. The model works by letting latent factor scores evolve through a hidden Markov chain, with each cluster contributing a Gaussian distribution for the factor scores. The authors derive a Gibbs sampler for posterior inference and show on a simulated dataset that credible intervals track the true parameter values and that cluster assignments are almost perfectly recovered. The payoff is a practical tool for longitudinal health data, illustrated here on recovery from opioid use disorder, where the estimated subgroups and transition probabilities describe different recovery trajectories.

What carries the argument

The central object is the hidden Markov chain on latent factor scores, combined with a factor model. The factor equation $y_{it}=B x_{it}+\epsilon_{it}$ ties observed variables to $L$ latent factors, while the hidden Markov structure on $Z_{it}$ lets the distribution of $x_{it}$ switch among Gaussian clusters $N(\mu_g,\Omega_g)$ over time. Identifiability is secured by a lower-triangular loadings matrix with unit diagonal and by requiring the first cluster's covariance to be diagonal. The Gibbs sampler alternates between sampling latent factors, cluster means and covariances, loadings, uniquenesses, initial and transition probabilities, and cluster assignments; the discrete full conditional for $Z_{it}$ in Equation (12) is the step that couples the factor model to the Markov chain.

What would settle it

Re-run the simulation but split the data: compute the empirical Bayes cluster priors from a training subset of subjects and run the Gibbs sampler on the remaining subjects. If the 95% credible intervals miss more than about 5% of true values or the misclassification rate rises substantially, the reported calibration depends on the double use of the data. Alternatively, replace the empirical Bayes priors with a fully Bayesian hierarchical prior and compare coverage on the same simulated data.

Watch

Extended reading notes

Core claim

The central claim is that BDCFM lets one coherent Bayesian analysis perform factor analysis, clustering, and dynamic transition estimation for longitudinal data. Observed vectors $y_{it}$ load onto $L$ latent factors through $y_{it}=B x_{it}+\epsilon_{it}$, each subject's factor vector at time $t$ is drawn from a cluster-specific Gaussian $x_{it}\mid Z_{it}=g\sim N(\mu_g,\Omega_g)$, and cluster membership $Z_{it}$ follows a hidden Markov chain with transition matrix $Q$. The paper supplies full conditional distributions for all parameters and implements a Gibbs sampler, then reports that in a simulation with $S=200$ subjects, $R=20$ variables, $T=5$ time points, $G=4$ clusters, and $L=3$ factors, 1.9% of 95% credible intervals missed the true parameter values and the cluster misclassification rate was 0.1%. In the opioid use disorder application with 252 subjects and 4 time points, the model identifies four recovery subgroups, estimates factor loadings with narrow credible intervals, estimates transition probabilities, and produces per-subject, per-time cluster assignments.

Load-bearing premise

The weakest premise is that the prior for each cluster's mean and covariance can be estimated from the same dataset that is later used for posterior inference; if that double use hides overfitting, the reported coverage and low misclassification rate would not generalize to new data.

Editorial extensions

If this is right

  • For a dataset of repeated multivariate measurements, BDCFM returns posterior distributions for factor loadings, cluster means and covariances, initial and transition probabilities, and each subject's cluster path over time.
  • The simulation evidence suggests the procedure is well calibrated: only 1.9% of 95% credible intervals missed true parameter values on the simulated dataset.
  • Cluster recovery is nearly exact in the reported simulation, with a misclassification rate of 0.1% for $S=200$, $T=5$, $G=4$, and $L=3$.
  • The opioid use disorder application identifies four recovery subgroups, gives estimated transition probabilities with credible intervals, and completes the analysis in 6.2 minutes for 252 subjects and 15 variables across 4 time points.
  • The model output can be visualized as alluvial plots of subject transitions, giving a dynamic account of how individuals move between recovery subgroups.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, a natural test not reported here is to split the data: compute the empirical Bayes cluster priors from a training subset and run the Gibbs sampler on the held-out subjects, then compare coverage and misclassification rates to see whether the double use of the data inflates apparent performance.
  • Beyond the paper, replacing the empirical Bayes priors with a fully Bayesian hierarchical prior for the cluster parameters would provide a direct check on how much of the reported calibration depends on the data-derived prior specification.
  • Beyond the paper, the same modeling structure could be extended to include subject-level covariates in the transition probabilities, allowing researchers to ask which factors make movement between recovery subgroups more or less likely.
  • Beyond the paper, the OUD application suggests a concrete clinical use: posterior transition probabilities between subgroups could help identify patients at risk of moving into a more severe recovery state, although that predictive use would need prospective validation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces Bayesian Dynamic Clustering Factor Models (BDCFM), which combine factor models with hidden Markov models for multivariate longitudinal data, aiming to perform dimension reduction, clustering, and estimation of dynamic cluster transitions. The authors propose conditionally conjugate priors, develop a Gibbs sampler for posterior exploration, and evaluate the method on a simulated dataset and on an opioid use disorder recovery study. The abstract claims accurate parameter estimation, high clustering accuracy, and well-calibrated uncertainty.

Significance. If the proposed method works as claimed, it would offer a useful tool for longitudinal clustering with uncertainty quantification in a high-dimensional setting. The paper provides a complete model specification and a Gibbs sampler with full conditional derivations, and it includes an application to a real dataset. However, the correctness of the sampler is compromised by an omitted factor in the initial-state full conditional, and the empirical Bayes priors are estimated from the same data used for coverage evaluation. The simulation evidence is additionally limited by the absence of comparisons to alternative methods. These issues substantially weaken the paper's central claims in its current form.

major comments (3)
  1. [Section 3, Eq. (12)] The full conditional for the initial cluster assignment Z_i1 omits the initial probability p_g. According to the joint model in Section 2, the conditional must be proportional to p_g N(x_i1 | mu_g, Omega_g) q_{g,Z_i2}. Since p_g is a model parameter updated in Eq. (10), it is not constant across g and cannot be dropped. Consequently, the Gibbs sampler described in Steps 2-10 does not target the stated posterior, and the simulation results in Section 4 (the 1.9% credible interval miss rate and the 0.1% misclassification rate) do not validate the model as specified.
  2. [Appendix 7.1] The empirical Bayes prior specification derives cluster mean and covariance hyperparameters from the same dataset that is subsequently used for posterior inference and coverage evaluation. Specifically, factor analysis and k-means clustering are applied to all observations to set m_mu_g, S_Omega_g, and s^2_omega_l. The reported 95% credible interval coverage in Section 4 is therefore conditional on data-dependent priors and does not provide a valid frequentist calibration. The authors should either use a split-sample approach, adopt fully Bayesian priors, or clearly discuss the implications of this double use of the data.
  3. [Section 4] The simulation study uses only a single dataset with well-separated cluster means and no comparison against alternative methods such as standard hidden Markov models, mixtures of factor analyzers, or other dynamic clustering approaches. This provides limited support for the abstract's claim that the inferential approach 'works well.' Adding comparative benchmarks or at least a discussion of relative performance would strengthen the evidence.
minor comments (4)
  1. [Section 6] The paper states that an R package will be available on CRAN upon publication, but no code repository or link is provided in the current version, which hampers reproducibility.
  2. [Section 4] The MCMC details are incomplete: no convergence diagnostics such as trace plots or Gelman-Rubin statistics are reported, and the choice of burn-in and thinning is stated without justification.
  3. [Section 3, Eq. (12)] The full conditional for Z_it is given only for the interior time points t = 2,...,T-1; the boundary cases t=1 and t=T are described in the text but not displayed as explicit equations, which makes the presentation less clear.
  4. [Table 1] The subgroup sample sizes n reported in Table 1 sum to 1008, which is larger than the number of subjects S=252; the table would benefit from a note clarifying that these are subject-time observations rather than unique subjects.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the BDCFM derivation is self-contained; the empirical-Bayes prior and a t=1 full-conditional omission are correctness concerns, not reductions of predictions to inputs.

full rationale

The model is defined in Section 2 as a factor model with HMM-driven latent states (Eq. 1 and the Z_it specification), and all full conditionals are derived from that joint density in Appendix 7.2 rather than imported from the authors' prior work. The simulation evaluation in Section 4 compares posterior estimates to fixed generative truth values, and the reported 0.1% misclassification is an output of the Gibbs sampler, not an identity with the k-means initialization used for empirical-Bayes hyperparameters. Appendix 7.1 uses the same dataset to set cluster hyperparameters; this double use of the data can inflate apparent performance and is a legitimate robustness limitation, but it does not make any posterior quantity equal to a fitted input by construction, and the paper explicitly labels the procedure empirical Bayes rather than presenting it as an external prior. The self-citations (Craft et al. 2023; Shin et al. 2025) are motivational or set G and L for the illustrative OUD analysis; the Gibbs sampler and full conditionals do not depend on those citations. The most serious issue in the paper is not circularity: Equation (12) omits the initial-state probability p_g from the t=1 full conditional for Z_i1, while Equation (10) and the joint density in Section 2 include p_g, so the sampler as written may not target the stated posterior. That is a correctness/bug concern, not a self-referential reduction, and is therefore not scored here as circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The model introduces no new physical or mathematical entities beyond standard latent factors and cluster indicators. The key extra assumptions are the empirical Bayes priors that use the same data twice and the fixed numbers of clusters and factors.

free parameters (4)
  • Empirical Bayes cluster mean hyperparameters m_mu_g (g=1,...,G) = k-means centroids of preliminary factor scores from the same data
    Used as prior means for cluster mean vectors in Section 7.1; derived from the data being analyzed.
  • Empirical Bayes cluster covariance hyperparameters S_Omega_g and s2_omega_l = Sample covariance matrices and diagonal elements of transformed factor scores from the same data
    Used as inverse Wishart and inverse gamma prior scale parameters in Section 7.1; derived from the same data.
  • Number of clusters G and number of factors L = G=4, L=3 for OUD analysis
    Chosen from prior study (Craft et al., 2023) rather than estimated within the model; no method for selecting G and L is provided.
  • Prior hyperparameters n_sigma=2.2, n_sigma s2_sigma=0.1, n_tau=1, alpha=2 = Fixed constants
    Chosen by hand following Lopes and West (2004) and to avoid zero initial probabilities; not the central focus of the paper.
assumptions (4)
  • domain assumption Factor loadings matrix B is lower triangular with unit diagonal to identify the factor model.
    Invoked in Section 2.1, following Geweke and Zhou (1996) and Aguilar and West (1999).
  • domain assumption Latent factors x_it given cluster assignment follow multivariate Gaussian distributions.
    Model definition in Section 2.1; standard for factor models.
  • domain assumption Cluster assignments follow a homogeneous hidden Markov model with transition matrix Q.
    Model definition in Section 2.1.
  • ad hoc to paper Empirical Bayes priors estimated from the same data provide valid posterior inference.
    Appendix 7.1 estimates prior hyperparameters from the data; this is an unvalidated assumption that may inflate performance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bayesian Dynamic Clustering Factor Models." pith.science (2026). https://pith.science/paper/VKRCXJEQ

@misc{pith2026250521490,
  author       = {Pith},
  title        = {Pith review of: Bayesian Dynamic Clustering Factor Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VKRCXJEQ}},
  note         = {Machine review of arXiv:2505.21490}
}
read the original abstract

We propose novel Bayesian Dynamic Clustering Factor Models (BDCFM) for the analysis of multivariate longitudinal data. BDCFM combines factor models with hidden Markov models to concomitantly perform dimension reduction, clustering, and estimation of the dynamic transitions of subjects through clusters. We develop an efficient Gibbs sampler for exploration of the posterior distribution. An analysis of a simulated dataset shows that our inferential approach works well both at parameter estimation and clustering of subjects. Finally, we illustrate the utility of our BDCFM with an analysis of a dataset on opioid use disorder.

Figures

Figures reproduced from arXiv: 2505.21490 by the authors.

Figure 1
Figure 1. Estimated factor loading matrix B using BDCFM with 4 subgroups and 3 factors on the OUD recovery data. Posterior means are indicated with dots and 95% credible intervals are indicated with vertical lines. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Estimated posterior density of µ using BDCFM with 4 subgroups and 3 factors on the OUD recovery data. Each panel represents the estimated posterior means of each cluster for one factor. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Estimated transition probability matrix Q using BDCFM with 4 subgroups and 3 factors on the OUD recovery data. (a) Heatmap of the posterior mean of transi￾tion probabilities. (b) Posterior mean and 95% credible interval of subgroup transition probabilities. On the x-axis, the two digit notation represents the subgroup transitions. Specifically, the first digit is the subgroup the subject is transitioning from, and t… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Subgroup assignment probabilities using BDCFM with 4 subgroups and 3 fac [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: Posterior modes of Zit using BDCFM with 4 subgroups and 3 factors on the OUD recovery data. The alluvial plot indicates the subgroup assignment at each time point and the subject-level transitions among subgroups over time. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 27 canonical work pages

  1. [1]

    Craft, Hwasoo Shin, Allison N

    William H. Craft, Hwasoo Shin, Allison N. Tegge, Diana R. Keith, Liqa N. Athamneh, Jeffrey S. Stein, Marco A. R. Ferreira, Howard D. Chilcoat, Anne Le Moigne, Angela DeVeaugh-Geiss, and Warren K. Bickel. Long-term recovery from opioid use disorder: recovery subgroups, transition states and their association with substance use, treatment and quality of lif...

    work page doi:10.1111/adb.13291 2023

  2. [2]

    Ka Yee Yeung and Walter L. Ruzzo. Principal component analysis for clustering gene expression data. Bioinformatics, 17 0 (9): 0 763--774, 2001

    work page 2001

  3. [3]

    The elements of statistical learning: data mining, inference, and prediction

    Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The elements of statistical learning: data mining, inference, and prediction. Springer, second edition, 2009

    work page 2009

  4. [4]

    Lopes and M

    H. Lopes and M. West. Bayesian model assessment in factor analysis. Statistica Sinica, 14: 0 41--67, 2004

    work page 2004

  5. [5]

    Ferreira

    Hwasoo Shin and Marco A.R. Ferreira. Dynamic ICAR spatiotemporal F actor M odels. Spatial Statistics, 56: 0 100763, 2023. doi:10.1016/j.spasta.2023.100763. URL https://doi.org/10.1016/j.spasta.2023.100763

    arXiv 2023

  6. [6]

    Fr\"uhwirth-Schnatter

    S. Fr\"uhwirth-Schnatter . Finite Mixture and M arkov Switching Models . Springer-Verlag, 2006

    work page 2006

  7. [7]

    Handbook of Mixture Analysis

    Sylvia Fruhwirth-Schnatter, Gilles Celeux, and Christian P Robert. Handbook of Mixture Analysis. CRC press, 2019

    work page 2019

  8. [8]

    Hwasoo Shin, Marco A. R. Ferreira, and Allison N. Tegge. Bayesian clustering factor models, 2025. URL https://arxiv.org/abs/2505.05280

    work page Pith review arXiv 2025

Show all 31 references
  1. [9]

    Sampling-based approaches to calculating marginal densities

    Alan E Gelfand and Adrian F M Smith. Sampling-based approaches to calculating marginal densities. Journal of the American Statistical Association, 85 0 (410): 0 398--409, 1990 a

    1990

  2. [10]

    C. P. Robert and G. Casella. Monte C arlo Statistical Methods . Springer-Verlag, New York, 2nd edition, 2005

    2005

  3. [11]

    Dani Gamerman and Hedibert F. Lopes. Markov Chain M onte C arlo: Stochastic Simulation for B ayesian Inference . Chapman and Hall/CRC, Boca Raton, FL, 2nd edition, 2006

    2006

  4. [12]

    A reference B ayesian test for nested hypotheses and its relationship to the S chwarz criterion

    Robert E Kass and Larry Wasserman. A reference B ayesian test for nested hypotheses and its relationship to the S chwarz criterion. Journal of the American Statistical Association, 90 0 (431): 0 928--934, 1995

    1995

  5. [13]

    Performance of B ayesian model selection criteria for G aussian mixture models

    Russell J Steele and Adrian E Raftery. Performance of B ayesian model selection criteria for G aussian mixture models. In M. Chen, Peter Muller, Dongchu Sun, Keying Ye, and Dipak Dey, editors, Frontiers of Statistical Decision Making and Bayesian Analysis, pages 113--130. Spri...

    2010

  6. [14]

    Mixtures of factor analysers

    Ernest Fokou \'e and DM Titterington. Mixtures of factor analysers. bayesian estimation and inference by stochastic simulation. Machine Learning, 50: 0 73--94, 2003

    2003

  7. [15]

    Overfitting B ayesian mixtures of factor analyzers with an unknown number of components

    Panagiotis Papastamoulis. Overfitting B ayesian mixtures of factor analyzers with an unknown number of components. Computational Statistics & Data Analysis, 124: 0 220--234, 2018

    2018

  8. [16]

    Escaping the curse of dimensionality in B ayesian model-based clustering

    Noirrit Kiran Chandra, Antonio Canale, and David B Dunson. Escaping the curse of dimensionality in B ayesian model-based clustering. Journal of Machine Learning Research, 24 0 (144): 0 1--42, 2023

    2023

  9. [17]

    Bayesian clustering of high-dimensional data via latent repulsive mixtures

    Lorenzo Ghilotti, Mario Beraha, and Alessandra Guglielmi. Bayesian clustering of high-dimensional data via latent repulsive mixtures. Biometrika, 112 0 (2): 0 asae059, 2025

    2025

  10. [18]

    Measuring the price of the arbitrage pricing theory

    John Geweke and Guofu Zhou. Measuring the price of the arbitrage pricing theory. The Review of Financial Studies, 9 0 (2): 0 557--587, 1996. ISSN 08939454, 14657368

    1996

  11. [19]

    Bayesian dynamic factor models and portfolio allocation

    Omar Aguilar and Mike West. Bayesian dynamic factor models and portfolio allocation. Journal of Business & Economic Statistics, 18 0 (3): 0 338--357, 1999. doi:10.1080/07350015.2000.10524868

    1999

  12. [20]

    Prado, M

    R. Prado, M. A. R. Ferreira, and M. West. Time Series: Modeling, Computation, and Inference. S econd Edition. Boca Raton: CRC Press, 2021

    2021

  13. [21]

    Gelfand and Adrian F

    Alan E. Gelfand and Adrian F. M. Smith. Sampling-based approaches to calculating marginal densities. Journal of the American Statistical Association, 85 0 (410): 0 398--409, 1990 b . doi:10.2307/2289776

    1990 doi

  14. [22]

    Sheehan, Yves Lecrubier, Kathleen Harnett Sheehan, Pedro Amorim, Judith Janavs, Edouard Weiller, T

    David V. Sheehan, Yves Lecrubier, Kathleen Harnett Sheehan, Pedro Amorim, Judith Janavs, Edouard Weiller, T. Hergueta, R. Baker, and Graeme C. Dunbar. The mini-international neuropsychiatric interview (mini): The development and validation of a structured diagnostic psychiatri...

    1998

  15. [23]

    Kessler, Patricia R

    Ronald C. Kessler, Patricia R. Barker, Lisa J. Colpe, Joan F. Epstein, Joseph C. Gfroerer, Eva Hiripi, Mary J. Howes, Sharon-Lise T. Normand, Ronald W. Manderscheid, Ellen E. Walters, and Alan M. Zaslavsky. Screening for serious mental illness in the general population. Archiv...

    2003 doi

  16. [24]

    Cochrane, Marvin J

    Lester Handelsman, Kenneth J. Cochrane, Marvin J. Aronson, Ronald Ness, Kenneth J. Rubinstein, and Philip D. Kanof. Two new rating scales for opiate withdrawal. American Journal of Drug and Alcohol Abuse, 13: 0 293--308, 1987. doi:10.3109/00952998709001515

    1987 doi

  17. [25]

    Reliability and validity of a modified brief pain inventory short form in patients with osteoarthritis

    Tito Mendoza, Tracy Mayne, Diane Rublee, and Charles Cleeland. Reliability and validity of a modified brief pain inventory short form in patients with osteoarthritis. European Journal of Pain, 10: 0 353--356, 2006. doi:10.1016/j.ejpain.2005.06.002

    2006 doi

  18. [26]

    Beck, Robert A

    Aaron T. Beck, Robert A. Steer, and Gregory K. Brown. Manual for the Beck Depression Inventory-II. Psychological Corporation, San Antonio, TX, 1996

    1996

  19. [27]

    John Jr Ware, Mark Kosinski, and Susan D. Keller. A 12-item short-form health survey: construction of scales and preliminary tests of reliability and validity. Medical Care, 34 0 (3): 0 220--233, 1996. doi:10.1097/00005650-199603000-00003

    1996 doi

  20. [28]

    W. Revelle. psych: Procedures for Psychological, Psychometric, and Personality Research. Northwestern University, Evanston, Illinois, 2024. URL https://CRAN.R-project.org/package=psych. R package version 2.4.3

    2024

  21. [29]

    G. H. Thomson. The Factorial Analysis of Human Ability. F ifth E dition. University of London Press, 1951

    1951

  22. [30]

    Estimating the number of clusters in a data set via the gap statistic

    Robert Tibshirani, Guenther Walther, and Trevor Hastie. Estimating the number of clusters in a data set via the gap statistic. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 63 0 (2): 0 411--423, 2001. doi:10.1111/1467-9868.00293

    2001

  23. [31]

    Trefethen and David Bau III

    Lloyd N. Trefethen and David Bau III. Numerical Linear Algebra. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1997. ISBN 978-0-898713-61-9

    1997

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.