Post-hoc mixture models of the best linear unbiased predictors (BLUP) from linear mixed-effects models: An approach for clustering unbalanced longitudinal data with irregularly spaced intervals

  • June 3, 2021 from 12:00-1:00pm
  • Please RSVP by emailing Lisa Maturo at
  • Please include your full name, email address, and institution/organization. 
  • We will provide instructions on obtaining CME credit for attendance.

Classification of longitudinal trajectories in naturally occurring data is an area of growing interest in learning healthcare systems and precision medical decisions.  While few approaches such as Group-based clustering (implemented in SAS proc Traj), an extension of k-means clustering (implemented in R package kml), and recently introduced deep learning methods (implemented in R, Python, and MATLAB) have shown great utility for clustering longitudinal data with fixed time, only mixture-based mixed-effects models, implemented in R package lcmm and M-plus, are mainly used for longitudinal data with irregularly spaced intervals. While theoretically sound, approaches of mixture-based mixed-effects models have appeared to be inadequate to fully address the classification problem of longitudinal unbalanced data. The approaches become computationally complex with increased data size and level of unbalancedness. We have shown that the application of post-hoc Gaussian finite mixtures on empirical BLUPs can identify the presence of subgroups. The method assumes a mixture of subgroups of normally distributed random effects with varying geometric features of the covariance matrix that can be determined by eigenvalue decomposition. In this presentation, we will introduce user-friendly post-hoc mixture modeling of BLUPs (PMMB) from mixed-effects models as an approach for clustering longitudinal data with haphazardly spaced intervals and will use real and simulated data to compare the classification performance of this method with that of existing mixture-based mixed-effects models.

Meet the Speaker

Jobayer Hossain, PhD

Dr. Jobayer Hossain is the Director of the Nemours Biostatistics Program and a Senior Research Scientist in the Department of Biomedical Research at Nemours Children’s Health System. My research interests focus on the innovative application of statistical methods to ensure effectiveness in studies involving clinical, epidemiological, and public health research. Identifying unobservable subgroups based on longitudinal changes in outcomes or cross-sectional complex associations is my current key interest besides modeling longitudinal and cross-sectional data for known groups. As the head of the Nemours Biostatistics Program, I participate in a wide variety of funded and non-funded research studies and provide statistical support to clinical researchers in the design of clinical trials, pre-clinical, epidemiological, and survey studies. This includes the provision of power and sample size calculations, randomization, data management, and data analysis; as well as authorship of the statistical section of research protocols, manuscripts, abstracts, and presentations. Commonly used statistical techniques include linear and nonlinear regressions with and without random coefficients (linear/generalized linear mixed-effects models, generalized estimating equations), multivariate methods (principal component, factor, discriminant, and cluster analyses, mixture models), latent growth mixture modeling (growth mixture model (GMM), heterogeneous linear mixed-effects model (HLME)), survival analysis, time series analysis, propensity score, bootstrap, cross-validation, G-estimation, and other parametric and nonparametric statistical techniques. In applications on the real and simulated datasets, my recently introduced method- “posthoc mixture modeling of BLUPs from mixed-effects models” showed better performance than widely used GMM and HLME in classifying trajectories in longitudinal unbalanced data. Major areas of my clinical interests include pediatric obesity, allergy, asthma, endocrine disorders, diabetes, cancer, and neurologic disorders. Besides my role as a co-investigator, I have led various funded and non-funded epidemiological studies as the principal investigator. I teach graduate and undergraduate statistics courses at the University of Delaware and provide training to research fellows at Nemours. I am a lead or co-author for more than 120 published articles and 100 published abstracts. I possess the expertise, leadership, training, experience, and motivation necessary for effective contribution to this study.

This activity has been approved for AMA PRA Category 1 Credit

Stay Connected

Join the ACCEL newsletter and keep up to date with the latest news, events and funding opportunities!