Mixed Models for Longitudinal Data Analysis

Mixed models for longitudinal data analysis provide a unified framework to examine repeated measurements collected over time from individuals or experimental units. By incorporating both fixed effects, which capture population‐level trends, and random effects, which account for subject‐specific deviations, these models enable robust estimation of trajectories while accommodating complex correlation structures. Longitudinal mixed models can handle unbalanced designs, unequal follow‐up times and missing data under plausible assumptions, making them indispensable in clinical trials, epidemiology, education and social sciences. Methodological advances have extended the classical linear mixed‐effects model to generalized linear formulations for non‐Gaussian outcomes, semiparametric approaches employing spline‐based smoothing for nonlinear time effects, and fully Bayesian frameworks allowing flexible distributional assumptions, such as skew‐t random effects. Modern computational techniques, including integrated nested Laplace approximations and Hamiltonian Monte Carlo algorithms, have expanded the feasibility of fitting high‐dimensional and nonlinear mixed models. Applications range from monitoring biomarker dynamics and cognitive development to evaluating treatment effects and identifying latent subgroups with distinct longitudinal patterns. The integration of mixed models with machine‐learning tools and high‐throughput data is forging new avenues for personalised prediction and adaptive study designs, underscoring the global significance and practical utility of this versatile statistical paradigm.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Researchers have developed a semiparametric Bayesian mixed‐effects model that combines spline‐based time functions with a multivariate skew‐t distribution for both random effects and residuals. This framework offers enhanced flexibility over traditional linear models, capturing asymmetry and heavy tails in biomarker trajectories and demonstrating superior fit in chronic disease studies.

Another line of work introduces fractional polynomial models tailored to non‐Gaussian longitudinal outcomes. By allowing time effects to follow power transformations within a generalized linear mixed‐effects setting, this approach adapts to binary and count data, revealing nuanced growth patterns in psychological and epidemiological studies and outperforming purely parametric specifications.

A recent bivariate semiparametric random‐coefficients model addresses multivariate longitudinal outcomes by estimating discrete support points for random effects via an EM algorithm. Applied to educational achievement data, this method uncovers latent classes of subjects with correlated trajectories across reading and mathematics, providing a rich basis for exploring contextual influences on learning over time.

Mixed Models for Longitudinal Data Analysis publication trend

The graph below shows the total number of articles in mixed models for longitudinal data analysis across all publications each year (not limited to Nature Index journals).

Technical terms

Mixed effects model: A statistical model that includes both fixed effects for population parameters and random effects for individual‐specific deviations.

Longitudinal data: Repeated measurements collected on the same experimental units or subjects over time.

Fixed effects: Parameters in a model assumed to be identical across all subjects, representing average trends or treatment impacts.

Random effects: Subject‐ or cluster‐specific parameters treated as random variables to capture individual variability and correlation.

Generalized linear mixed model: An extension of mixed models that allows non‐Gaussian response distributions via link functions.

Semiparametric model: A modeling approach combining parametric components with nonparametric elements, such as spline‐based functions for flexibility.

Skew‐t distribution: A flexible distribution generalising the normal to accommodate skewness and heavy tails in random effects or residuals.

Integrated nested Laplace approximation (INLA): A computational method for fast approximate Bayesian inference in latent Gaussian models, including mixed effects.

References

1. Flexible Bayesian semiparametric mixed-effects model for skewed longitudinal data. BMC Medical Research Methodology (2024).
2. Fitting the Fractional Polynomial Model to Non-Gaussian Longitudinal Data. Frontiers in Psychology (2017).
3. Evaluating class and school effects on the joint student achievements in different subjects: a bivariate semiparametric model with random coefficients. Computational Statistics (2021).
4. Improving the INLA approach for approximate Bayesian inference for latent Gaussian models. Electronic Journal of Statistics (2015).