📊 statistics

Meta-analysis with the glmmTMB R package

This paper introduces the `equalto` covariance structure in the R package `glmmTMB`, enabling flexible meta-analytical modeling with known sampling variances that matches the accuracy of the established `metafor` package while offering enhanced capabilities for handling heteroscedasticity and dependent errors.

Original authors: Coralie Williams, Maeve McGillycuddy, Mollie Brooks, Benjamin M. Bolker, Ayumi Mizuno, Yefeng Yang, Wolfgang Viechtbauer, David I. Warton, Shinichi Nakagawa

Published 2026-04-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Coralie Williams, Maeve McGillycuddy, Mollie Brooks, Benjamin M. Bolker, Ayumi Mizuno, Yefeng Yang, Wolfgang Viechtbauer, David I. Warton, Shinichi Nakagawa

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 detective trying to solve a mystery. You have collected clues from 50 different witnesses. Each witness saw a slightly different version of the event, and each has a different level of reliability (some have bad eyesight, some were far away, some were very close).

Meta-analysis is the process of combining all these witness accounts to figure out the "true" story.

For a long time, statisticians had a very specific, specialized tool for this job (like a dedicated detective's magnifying glass called the metafor package in R). It worked great, but it was a bit rigid. If you wanted to look at the clues in a weird new way, or if the clues were messy (like counting rare events instead of measuring heights), that specialized tool sometimes struggled.

Then, there was a Swiss Army Knife called glmmTMB. It was a general-purpose tool used by scientists to solve all kinds of complex puzzles. It was incredibly fast and flexible, but it had a missing blade: it couldn't easily handle the "known reliability" of the witnesses (the sampling variances) in the way meta-analysis requires.

This paper introduces a brand new blade for that Swiss Army Knife.

Here is the breakdown of what they did, using simple analogies:

1. The Problem: The "Known Error" Gap

In meta-analysis, we don't just guess how reliable a study is; we know the math behind its error.

The Old Way: You had to use the specialized detective tool (metafor) to handle these known errors.
The Problem: If you wanted to do something fancy—like account for the fact that some studies have multiple related clues, or that the data isn't a simple number but a count of rare animals—the specialized tool got stuck.

2. The Solution: The "Equalto" Blade

The authors (a team of statisticians) added a new feature to the Swiss Army Knife (glmmTMB) called equalto.

Think of equalto as a custom-fit adapter.

Normally, the Swiss Army Knife assumes every clue has a standard, random amount of fuzziness.
With the equalto adapter, you can say: "Hey, for this specific clue, the fuzziness is exactly equal to this specific number I calculated."

This allows the general-purpose tool to perfectly mimic the specialized detective tool, but with the added superpowers of the Swiss Army Knife.

3. What Superpowers Does This Unlock?

Once you have this adapter, you can solve puzzles that were previously impossible or very hard:

The "Family Tree" Puzzle (Phylogenetics): Imagine studying how different species of birds behave. Birds that are closely related (like cousins) share more history, so their data isn't independent. The new tool can easily map out these family trees and adjust the math accordingly.
The "Messy Count" Puzzle: Sometimes you aren't measuring a height; you are counting how many times a rare event happened (like a shark attack). This data is "spiky" and weird. The new tool can handle these "spiky" counts directly without forcing them into a square peg hole.
The "Double Trouble" Puzzle (Bivariate): Sometimes a study gives you two related results at once (e.g., how a drug affects both blood pressure and heart rate). The new tool can juggle both results simultaneously, understanding how they influence each other.
The "Variable Chaos" Puzzle (Location-Scale): Usually, we assume all studies are equally messy. But what if big studies are very precise, while small studies are chaotic? The new tool can let the "messiness" change depending on the size of the study.

4. Did It Work?

The authors ran a massive simulation (a "stress test").

They fed the same data into the old specialized tool and the new Swiss Army Knife with the adapter.
The Result: The answers were identical. The new tool gave the exact same "true story" as the old tool, but it was often faster when the puzzle got complicated.

The Bottom Line

Before this paper, if you wanted to do a simple meta-analysis, you used one tool. If you wanted to do a complex, messy, multi-layered meta-analysis, you might have been stuck or had to use very slow, complicated methods.

Now, researchers can use one single, powerful, flexible tool (glmmTMB) to do everything. It's like upgrading from a basic screwdriver to a high-tech, multi-tool that can handle any screw, bolt, or wire you throw at it, while still being just as accurate as the specialized tools you used to rely on.

In short: They made the general-purpose math tool smart enough to handle the specific rules of meta-analysis, opening the door to more complex and accurate scientific discoveries.

1. Problem Statement

Meta-analysis is a cornerstone of evidence synthesis, typically formulated as a mixed-effects model where sampling variances of effect sizes are treated as known. While specialized R packages like metafor are the standard for fitting these models, they lack the flexibility of general-purpose Generalized Linear Mixed Model (GLMM) software.

Limitation of Existing Tools: Standard GLMM packages (e.g., lme4) cannot easily incorporate known sampling error variance-covariance matrices, which are essential for handling heteroscedasticity and dependence among effect sizes (e.g., multiple outcomes per study).
The Gap: Researchers often need to fit complex meta-analytic models involving non-Gaussian outcomes (binary, count data), multiple random effects, dispersion modeling, or zero-inflation. Existing specialized meta-analysis tools often struggle with these complex structures, while general GLMM tools lack the mechanism to fix sampling variances to known matrices.

2. Methodology

The authors introduce a new covariance structure, equalto, within the glmmTMB R package. This structure allows users to supply a known sampling error variance-covariance matrix directly into the random effects specification.

Core Technical Implementation

The equalto() Function: This new covariance structure fixes the variance-covariance matrix of a specific random effect to a user-supplied matrix ( $V$ $V$ ).
- Syntax: equalto(0 + id|g, V)
  - id: A factor indexing the effect sizes.
  - g: A grouping factor (typically a single level for the whole dataset) defining independent random-effect vectors.
  - V: A symmetric matrix containing known sampling variances on the diagonal and covariances on the off-diagonals.
Model Frameworks Supported:
1. Conventional Meta-Analysis (Two-Stage): Fits the Normal-Normal model ( $y_i = \mu + u_i + e_i$ ) where $e_i$ has a known variance $v_i$ . The equalto structure replaces the residual error term with the fixed sampling variance matrix.
2. Multilevel/Hierarchical Meta-Analysis: Handles dependent effect sizes (e.g., multiple outcomes per study) by modeling the sampling error covariance matrix ( $V$ ) as a block-diagonal matrix with off-diagonal correlations.
3. GLMM Meta-Analysis (One-Stage): Extends the framework to model raw binary or count data directly using Binomial-Normal or Poisson-Normal models, avoiding the need to calculate standardized effect sizes beforehand.
Additional Flexibility: The implementation supports:
- Dispersion Modeling (dispformula): Allowing the residual variance (within-study heterogeneity) to vary based on moderators.
- Zero-Inflation (ziformula): Handling excess zeros in count data meta-analyses.
- Complex Random Effects: Integration with other covariance structures (e.g., propto for phylogenetic trees, ar1 for temporal data).

3. Key Contributions

Bridging the Gap: Provides a unified framework where standard mixed-model syntax (familiar to lme4 users) can be used for rigorous meta-analysis.
Enhanced Flexibility: Enables the fitting of meta-analytic models that are currently unsupported or difficult in specialized packages, such as:
- Models with multiple random effects in the dispersion formula.
- Non-linear smooth functions for covariates.
- Phylogenetic and spatial meta-analyses within a GLMM framework.
Open-Source Expansion: Expands the R toolkit for evidence synthesis by leveraging the computational efficiency of the Template Model Builder (TMB) and Laplace approximation used in glmmTMB.

4. Results

The authors validated the new implementation through simulation studies and re-analysis of published datasets.

Simulation Study:
- Accuracy: The glmmTMB estimates (overall mean $\hat{\mu}$ , between-study variance $\hat{\tau}^2$ , and standard errors) were identical to those produced by the established metafor package (specifically rma.uni and rma.mv) up to the 5th or 6th decimal place across various effect size measures (SMD, log response ratio, odds ratio, incidence ratio).
- Performance:
  - For simple random-effects models, metafor was slightly faster (~0.5s).
  - For complex models (e.g., phylogenetic meta-analysis with multiple random effects), glmmTMB was significantly faster (approx. 3x faster in the phylogenetic example: 19s vs. 65s).
Illustrative Applications:
- Bivariate Meta-Analysis: Successfully modeled treatment-specific means and covariances for BCG vaccine efficacy data.
- Phylogenetic Meta-Analysis: Integrated phylogenetic correlation matrices (propto) alongside sampling error structures (equalto) for plant diversity data, demonstrating superior speed over metafor.
- Location-Scale Meta-Analysis: Modeled heterogeneity as a function of sample size using the dispformula, a feature not yet supported in metafor's rma function.

5. Significance

Methodological Advancement: The equalto structure democratizes access to complex meta-analytic modeling. Researchers familiar with lme4 and glmmTMB can now perform evidence synthesis without switching to specialized, often less flexible, software.
Handling Complex Data: It addresses critical limitations in current tools by allowing the modeling of:
- Dependence: Explicitly modeling correlations between sampling errors (e.g., shared controls).
- Non-Gaussian Outcomes: Direct analysis of raw binary/count data without transformation.
- Heteroscedasticity: Modeling how variance changes across studies (location-scale models).
Future Directions: While glmmTMB currently lacks some post-hoc features found in metafor (e.g., specific subgroup plotting tools), the authors suggest that future work could integrate visualization tools (like orchard plots) with glmmTMB output. The implementation also opens the door for exploring alternative estimation methods (e.g., Gauss-Hermite quadrature) for GLMM meta-analyses.

In conclusion, the paper presents a robust, flexible, and computationally efficient extension to the glmmTMB package, significantly enhancing the capabilities of the R ecosystem for modern, complex evidence synthesis.