📊 statistics

ProfileGLMM: a R Package Extending Bayesian Profile Regression using Generalised Linear Mixed Models

The paper introduces ProfileGLMM, an R package that extends Bayesian profile regression by integrating Generalised Linear Mixed Models to simultaneously explain outcome variation and cluster observations, thereby enabling the analysis of hierarchical and longitudinal data with complex, correlated covariates.

Original authors: Matteo Amestoy, Mark A. van de Wiel, Wessel N. van Wieringen

Published 2026-04-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Matteo Amestoy, Mark A. van de Wiel, Wessel N. van Wieringen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to figure out why some patients are getting sick while others stay healthy. You have a huge pile of data: their age, where they live, what they eat, their job, and their medical history.

The Problem:
If you try to look at all these factors at once using a standard "recipe" (a normal statistical model), you get confused. Why? Because many of these factors are tangled together. For example, people who live in cities might also have different jobs and eat different foods. It's like trying to untangle a giant knot of headphones while wearing boxing gloves. Standard models often get stuck or give you a messy, inaccurate answer.

The Old Solution (Bayesian Profile Regression):
Previously, statisticians had a clever trick called "Profile Regression." Instead of looking at every single factor separately, they grouped patients into "profiles" or "neighborhoods" based on their similarities.

Analogy: Imagine sorting a messy box of LEGO bricks not by color or size individually, but by finding bricks that naturally fit together to build a specific shape. Once you've built the shapes (the clusters), you just ask: "Do the people in the 'City Apartment' shape get sick more often than the 'Suburban House' shape?"

The New Solution (ProfileGLMM):
The paper introduces a new tool called ProfileGLMM. It takes that clever LEGO-sorting trick and supercharges it with two major upgrades:

1. The "Time Travel" Upgrade (Handling Repeated Data)

The Old Way: The old method could only look at a snapshot in time. It was like taking one photo of a patient and guessing their health.
The New Way: ProfileGLMM can handle longitudinal data. This means it can track the same person over time (like a video instead of a photo).

Analogy: Imagine you are watching a movie of a patient's life. You see them at age 20, 30, and 40. The old method treated these as three different strangers. ProfileGLMM knows, "Hey, that's the same person!" It accounts for the fact that your health at 30 is influenced by your health at 20. It separates the "person's unique story" from the "general trends," giving a much clearer picture.

2. The "Secret Ingredient" Upgrade (Interactions)

The Old Way: The old method assumed that once you put someone in a "profile," that was the end of the story. It didn't ask how that profile interacted with other specific factors.
The New Way: ProfileGLMM allows for interactions. It can ask, "Does being in the 'City Apartment' profile make the effect of smoking worse?"

Analogy: Think of a cooking show. The old method said, "This group of people likes spicy food." The new method says, "This group of people likes spicy food, but only if they are also drinking coffee." It finds the hidden recipes that only work when specific ingredients are mixed together.

How It Works (The Magic Machine)

The package uses a computer algorithm (a Gibbs Sampler) that acts like a very smart, tireless detective.

It guesses: It randomly groups people into different "profiles."
It checks: It sees if these groups make sense with the health outcomes.
It refines: It shuffles people around, trying new groups, over and over again (thousands of times).
It settles: Eventually, it finds the most stable, logical way to group everyone.

Once it has the groups, it builds a custom "health report" for each group, explaining exactly how different factors (like pollution or diet) affect them, while also remembering that some people just have unique, personal health histories.

Why Should You Care?

This tool is a game-changer for researchers in medicine, sociology, and environmental science.

For Doctors: It helps find specific subgroups of patients who might need special treatments, rather than a "one size fits all" approach.
For Scientists: It allows them to analyze complex, messy real-world data (like tracking air pollution effects on the same people over years) without getting lost in the noise.

In a Nutshell:
ProfileGLMM is like giving a detective a time machine and a magnifying glass. It doesn't just sort people into groups; it watches how those groups change over time and discovers the secret recipes that link their lifestyles to their health outcomes. It turns a tangled knot of data into a clear, actionable story.

1. Problem Statement

Bayesian profile regression is a powerful statistical framework designed to simultaneously model an outcome variable and cluster observations based on a set of interdependent covariates (clustering covariates). While effective for cross-sectional data, existing implementations (e.g., the PReMiuM package) face two critical limitations when applied to complex modern datasets:

Inability to handle hierarchical/longitudinal structures: Standard profile regression assumes independent observations. It cannot account for within-subject correlations or unobserved heterogeneity found in repeated-measures or clustered data (e.g., longitudinal studies).
Lack of interaction modeling: Previous methods often treat latent clusters as static offsets, failing to model complex interactions between these latent clusters and other observable covariates.
Sub-optimal handling of correlated covariates: Directly including highly correlated clustering covariates in standard regression models leads to statistical inefficiency.

2. Methodology

The authors propose ProfileGLMM, an R package that integrates Generalised Linear Mixed Models (GLMMs) into the Bayesian profile regression framework.

Statistical Framework

The model decomposes the joint distribution of the outcome ( $y_i$ ) and clustering covariates ( $u_i$ ) conditional on regression covariates ( $x_i$ ) into a mixture model:
$p(y_i, u_i | \pi, \theta, x_i, \alpha) = \sum_{c=1}^{C} \pi_c f(y_i, u_i | \theta_c, x_i, \alpha)$

Key components include:

Assignment Distribution ( $f_u$ ): Models the clustering covariates ( $u_i$ ). It supports continuous variables (Multivariate Gaussian), categorical variables (Multinomial), or a mixture of both.
Outcome Distribution ( $f_y$ ): This is the novel contribution. Instead of a simple linear model, the outcome is modeled using a GLMM:
$y_i = x_{i}^{FE}\beta + x_{i}^{RE}\eta_{g(i)} + x_{i}^{Int}\gamma_c + \epsilon_i$
- Fixed Effects ( $\beta$ ): Standard regression coefficients.
- Random Effects ( $\eta$ ): Account for hierarchical structures (e.g., repeated measures within individuals).
- Cluster Interactions ( $\gamma_c$ ): Parameters specific to each latent cluster $c$ , allowing the effect of covariates to vary by cluster.
Infinite Mixture: The number of clusters ( $C$ ) is treated as infinite using a Dirichlet Process Mixture Model (DPMM) with a stick-breaking prior. To ensure computational tractability, a truncated Dirichlet process is employed, avoiding the need for complex Metropolis-within-Gibbs steps for cluster allocation.

Inference

Sampling: A Gibbs sampler (MCMC) is used to sample from the posterior distribution.
Post-Processing: To address the "label switching" problem inherent in MCMC mixture models, the package computes an average dissimilarity matrix to derive a representative clustering ( $Z^*$ ). Cluster-specific parameters are then estimated by aggregating posterior samples across the chain and observations assigned to the representative clusters.

3. Key Contributions

Integration of GLMMs: The primary innovation is embedding GLMMs as the outcome model within Bayesian profile regression. This allows for the analysis of longitudinal and hierarchical data by incorporating random effects.
Interaction Modeling: The framework explicitly models interactions between latent clusters and observable covariates ( $x_{i}^{Int}\gamma_c$ ), enabling the detection of heterogeneous effects across different exposure profiles.
Piecewise Linear Fitting: The package demonstrates a unique capability to treat a predictor as both a clustering covariate and an interaction term. This allows for non-parametric piecewise linear fitting, effectively segmenting data into regimes with distinct linear trends without pre-specifying break points.
Software Implementation: The ProfileGLMM R package is built on Rcpp for computational efficiency, offering a user-friendly interface with minimal mandatory parameters while supporting continuous and binary outcomes.

4. Results

The paper validates the methodology through two primary examples:

Simulated Longitudinal Exposome Data:
- A dataset of 1,500 individuals with 4,500 observations was generated with known latent clusters and random effects.
- Outcome: The model successfully recovered the nine true latent clusters, accurately estimated the fixed effects, random effects, and cluster-specific interaction terms. The representative clustering centroids closely matched the true underlying mesh grid.
Piecewise Linear Fit:
- A simple dataset with a known piecewise linear relationship between a predictor and an outcome was analyzed.
- Outcome: The model successfully identified six distinct local dynamics (segments) and estimated the specific linear trend for each segment, demonstrating its ability to model complex functional relationships non-parametrically.

Computational Performance:

Benchmarks against the established PReMiuM package showed that ProfileGLMM is computationally competitive.
For 12,000 observations, ProfileGLMM took approximately 485 seconds compared to 1,031 seconds for PReMiuM (under simplified comparable settings), indicating superior efficiency for large datasets.

5. Significance

Methodological Advancement: ProfileGLMM bridges the gap between profile regression and mixed modeling, enabling rigorous analysis of complex, correlated data structures common in epidemiology, social sciences, and clinical trials.
Practical Utility: By handling high-dimensional, correlated covariates and hierarchical data simultaneously, it provides a unified inference strategy that avoids the statistical sub-optimality of sequential approaches.
Flexibility: The ability to model interactions between latent clusters and covariates, as well as perform piecewise fitting, opens new avenues for exploratory data analysis where the functional form of relationships is unknown.
Accessibility: The package is freely available on CRAN, offering a robust, Rcpp-accelerated tool for researchers dealing with complex data structures, with future plans to include variable selection and time-to-event outcomes.