📊 statistics

Matrix Decomposition Latent Growth Model Tree

This paper introduces the Matrix Decomposition Latent Growth Model Tree (MDLGM Tree), a novel method that leverages matrix decomposition factor analysis to significantly reduce improper solutions and computational costs compared to conventional SEM Tree approaches when analyzing longitudinal data.

Original authors: Naoya Todo (Tokyo Metropolitan University), Naoto Yamashita (Kansai University), Satoshi Usami (The University of Tokyo)

Published 2026-08-05

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Naoya Todo (Tokyo Metropolitan University), Naoto Yamashita (Kansai University), Satoshi Usami (The University of Tokyo)

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 about how people change over time. In the world of psychology and behavioral science, researchers often look at "longitudinal data," which is just a fancy way of saying they track the same group of people over and over again—like taking a photo of a plant every week to see how it grows. To make sense of these changing patterns, scientists use a tool called a Latent Growth Model (LGM). Think of an LGM as a flexible ruler that draws a smooth line through a messy cloud of points to show the average path of growth, while also measuring how much each individual person's path wiggles away from that average.

But here's the twist: not everyone grows the same way. Some kids might learn math quickly at first and then slow down, while others might start slow and then zoom ahead. If you try to force everyone into one single "average" story, you miss the real differences. To fix this, researchers use a technique called a Structural Equation Model Tree (SEM Tree). Imagine this as a giant, digital decision tree. It asks questions like, "Do you have a high family income?" or "Are you a boy or a girl?" and then splits the group of people into smaller branches based on the answers. Each branch gets its own unique growth story. This is incredibly useful for finding hidden subgroups, but there's a catch: the current version of this digital tree is very heavy and clumsy. It takes a massive amount of computer power to build, and it often trips over its own feet, producing "improper solutions"—mathematical errors where the computer calculates impossible things, like a negative amount of growth or a broken map.

This is where the new study by Naoya Todo and his team steps in. They have invented a lighter, faster, and more reliable version of this digital tree called the Matrix Decomposition Latent Growth Model Tree (MDLGM Tree). Instead of using the heavy, traditional math that causes the computer to stumble, they use a clever trick called Matrix Decomposition Factor Analysis (MDFA). You can think of MDFA as a different way of solving a puzzle. While the old method tries to fit the pieces together by guessing and checking until the picture looks perfect (which often leads to errors), the new method breaks the picture down into its basic building blocks (like separating the colors from the shapes) and reconstructs it using a step-by-step process that mathematically guarantees the pieces will fit without breaking.

In a series of computer simulations, the authors tested their new tree against the old one. They found that the MDLGM Tree is a game-changer for speed; in some cases, it was more than 10 times faster than the traditional method, especially when tracking people over many time points (like 10 different measurements). More importantly, it almost completely eliminated those frustrating "improper solutions" where the math breaks down. While the old tree frequently produced errors (sometimes in over 50% of difficult cases), the new tree stayed stable. The study also showed that when the new tree did find a solution, the results were nearly identical to the old method, meaning it didn't sacrifice accuracy for speed. The authors suggest that this new tool could help researchers analyze complex growth data much more easily, though they note that more work is needed to make the software available to everyone and to test it on even more complex types of growth stories.

Technical Summary: Matrix Decomposition Latent Growth Model Tree

Problem Statement
Latent Growth Models (LGMs) are standard tools for analyzing individual trajectories of change and group differences in longitudinal data. When populations are heterogeneous, Structural Equation Model Trees (SEM Trees) offer a solution by recursively partitioning samples based on external covariates to identify subgroups with distinct growth parameters. However, current implementations of SEM Trees face two significant practical obstacles:

Computational Cost: Estimating the template model (LGM) via Maximum Likelihood (ML) at every candidate split is computationally demanding, particularly with large datasets or many time points.
Improper Solutions: ML estimation frequently yields improper solutions (e.g., negative residual variances, non-positive definite covariance matrices) when models are misspecified or when sample sizes within nodes are limited. Existing software packages (e.g., semtree, MplusTrees) often lack robust, built-in mechanisms to automatically exclude nodes solely based on the occurrence of such improper solutions, potentially leading to unstable tree structures or biased parameter estimates.

Methodology
The authors propose the Matrix Decomposition Latent Growth Model Tree (MDLGM Tree), a novel recursive partitioning method that replaces the standard ML estimation of the template LGM with a Matrix Decomposition Factor Analysis (MDFA) approach.

MDLGM Formulation: The method extends MDFA to the context of LGMs, accommodating mean structures and factor covariances. Unlike standard SEM which minimizes a likelihood-based loss function in covariance space, MDLGM treats factor scores and unique variances as fixed parameters and minimizes a Frobenius norm-based least squares criterion in the data space.
- The linear and quadratic LGMs are reformulated as confirmatory factor analysis models where factor loadings are fixed (e.g., $1, t-1$ for linear; $1, t-1, (t-1)^2$ for quadratic).
- The estimation procedure involves an iterative eigenvalue-decomposition algorithm. Crucially, the unique variance estimates are derived from the diagonal of a matrix product involving eigenvectors, which theoretically prevents negative variance estimates during the iterative update process.
- To handle local optima, the algorithm employs multiple random starts, selecting the solution with the minimum loss function.
Tree Construction Algorithm:
- Splitting Criterion: Instead of relying on log-likelihood or deviance changes (which require distributional assumptions), the MDLGM Tree uses the squared Mahalanobis distance between observed data and model-implied values as the splitting criterion.
- Procedure: For a parent node, the template model is fitted via MDLGM. The data is then split at potential cut-off points of covariates. The sum of Mahalanobis distances for the resulting child nodes is calculated. The split that maximizes the reduction in distance (subject to a complexity parameter, $cp$) is selected.
- Improper Solution Handling: Because the MDFA-based estimation inherently avoids negative variance estimates, the algorithm naturally mitigates the occurrence of improper solutions that typically halt or destabilize ML-based tree construction.

Key Contributions

Algorithmic Innovation: The development of MDLGM, which adapts matrix decomposition techniques to estimate LGM parameters including means and covariances, extending the scope of MDFA beyond standard factor analysis.
Integration with Decision Trees: The creation of the MDLGM Tree, which integrates this estimation method with recursive partitioning, offering a computationally efficient alternative to SEM Trees.
Robustness: The method provides a mechanism to substantially reduce the risk of improper solutions (negative variances, singular matrices) without requiring ad-hoc constraints or complex Bayesian priors.

Results
A series of simulation studies compared MDLGM Tree (two variants: Mahalanobis distance and deviance-based splitting) against four SEM Tree variants (ML, Constrained ML, Bayesian, and ML with warning-based exclusion) under various conditions (linear vs. quadratic models, correct vs. misspecified models, varying sample sizes $N=500, 2000$ , time points $T=5, 10$ , and separation degrees).

Computational Efficiency: MDLGM Tree was significantly faster than all ML-based and Bayesian SEM Tree methods. Under conditions with a large number of time points ( $T=10$ ), MDLGM Tree was, on average, more than 10 times faster. Its computation time was largely insensitive to increases in sample size or time points, unlike SEM Tree methods.
Improper Solutions: MDLGM Tree produced almost no improper solutions. In contrast, ML-based SEM Trees frequently encountered improper solutions (exceeding 50% in some quadratic model conditions), and even constrained ML or warning-exclusion strategies failed to eliminate them entirely. Bayesian SEM Trees showed no improper solutions but suffered from other issues (see below).
Node Estimation Accuracy: MDLGM Tree (specifically the Mahalanobis distance variant) consistently and accurately estimated the true number of terminal nodes across all conditions.
- Bayesian SEM Trees (SEM Tree3) performed poorly in estimating the number of nodes under most conditions, suggesting high sensitivity to prior specification.
- ML-based methods showed reduced accuracy in node estimation when $T$ increased, likely due to the difficulty of distinguishing nodes when model misspecification or specific parameter configurations obscured differences.
Parameter Estimates: When the number of nodes was correctly identified, the parameter estimates (means, variances, covariances) from MDLGM Tree were highly correlated with those from ML-based methods (correlations close to 1.0), indicating that MDLGM Tree yields solutions similar to ML while avoiding its computational and stability pitfalls.
Classification Accuracy: When nodes were correctly identified, classification accuracy was generally high across methods. However, SEM Tree4 (which excludes splits based on warnings) showed lower classification accuracy in some conditions, likely because it avoided splitting on relevant covariates due to improper solutions in parent nodes.

Significance and Claims
The paper claims that MDLGM Tree is a useful model-based recursive partitioning method for longitudinal data analysis. Its primary significance lies in its ability to:

Detect group differences in latent trajectories more reliably by avoiding the improper solutions that plague current SEM Tree implementations.
Identify external covariates that explain these differences with significantly lower computational cost, making it feasible to apply to larger datasets or models with more time points.
Classify individuals more accurately by ensuring that the underlying growth models in each node are properly estimated.

The authors position MDLGM Tree not only as a standalone tool but also as a potential instrument for sensitivity analysis alongside SEM Trees. They acknowledge limitations, noting that the current implementation focuses on simple linear and quadratic LGMs and that future work is needed to extend the algorithm to more complex LGM specifications (e.g., free factor loadings), investigate theoretical properties (consistency, efficiency), and develop general-purpose software packages. The paper concludes that by addressing the computational and stability barriers of SEM Trees, MDLGM Tree has the potential to promote the wider practical use of longitudinal SEM approaches in psychological and behavioral sciences.

More like this