📊 statistics

Shrinkage Estimators in Finite Mixture of Mixed Generalized Estimating Equations

This paper proposes an iterative shrinkage weighted least squares algorithm based on penalized likelihood to estimate parameters for finite mixture models of mixed generalized estimating equations, addressing population heterogeneity where standard linear mixed models fail, and validates the method's performance through Monte Carlo simulations.

Original authors: Sajjad Nezamdoust, Farzad Eskandari

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sajjad Nezamdoust, Farzad Eskandari

Original paper licensed under CC BY 4.0 (https://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 trying to understand a large, noisy crowd of people. In a standard statistical model, you might assume everyone in the crowd is roughly the same type of person, just with slight variations in height or weight. You'd draw one big average line to describe the whole group.

But what if the crowd is actually made up of several distinct subgroups? Maybe you have a group of basketball players, a group of jockeys, and a group of sumo wrestlers all mixed together. If you try to draw one single line to describe the "average" person in this crowd, you'll get a mess. You won't accurately describe the basketball players, the jockeys, or the sumo wrestlers. You need to realize the crowd is a mixture of different types.

This is the core problem this paper tackles. The authors, Sajjad Nezamdoust and Farzad Eskandari, are dealing with data that comes from multiple hidden groups (a "Finite Mixture") where each individual in those groups has their own unique, repeating patterns (a "Mixed Model").

Here is a simple breakdown of their solution, using everyday analogies:

1. The Problem: The "One-Size-Fits-All" Trap

The paper argues that when data comes from a population with hidden subgroups (heterogeneity), standard models fail. It's like trying to fit a single pair of shoes on a foot that is actually a collection of three different feet glued together. The model gets confused, and the predictions are poor.

2. The Solution: A "Smart" Sorting Hat

The authors propose a new way to estimate the numbers (parameters) that define these hidden groups. They call their method Shrinkage Estimators within a Finite Mixture of Mixed Generalized Estimating Equations.

Let's break that jargon down:

Finite Mixture: Acknowledging the crowd is made of distinct subgroups (like the basketball players vs. jockeys).
Mixed Model: Acknowledging that within those groups, individuals have their own personal "quirks" or random effects (like a specific player's bad knee affecting their height measurement).
Shrinkage: This is the magic trick. Imagine you are guessing the height of a person. If you just look at one measurement, you might be way off. "Shrinkage" is like taking your wild guess and gently pulling it closer to a sensible average. It prevents the model from getting too excited by random noise or weird outliers. It "shrinks" the extreme guesses toward a more realistic center.

3. The Tools: Three Different "Shrinking" Techniques

The authors tested three specific ways to do this "shrinking," comparing them like three different tools in a toolbox:

Ridge Regression: Think of this as a gentle, steady hand. It pulls all the guesses slightly toward the center, but it doesn't make any of them zero. It's like smoothing out a bumpy road so it's easier to drive on, but the road is still there.
Lasso Regression: This is a more aggressive tool. It doesn't just pull guesses toward the center; it can cut them off completely (making them zero). It's like a sculptor chiseling away the unnecessary parts of a statue. If a specific factor (like "stolen bases" in baseball) doesn't really matter for the salary, Lasso removes it entirely from the equation.
Elastic Net: This is the best of both worlds. It's a hybrid tool that combines the gentle smoothing of Ridge with the aggressive cutting of Lasso. It's like a Swiss Army knife that can both smooth and cut depending on what the data needs.

4. The Process: Iterative Weighted Least Squares (IWLS)

How do they actually find the answers? They use an algorithm called Iterative Weighted Least Squares.

Imagine you are trying to balance a stack of books on a wobbly table.

You place the books (make a guess).
The table wobbles (the data is noisy).
You adjust the books based on how they fell (calculate the error).
You place them again, but this time you weigh the books differently based on how stable they were last time.
You repeat this process over and over until the stack is perfectly balanced and doesn't wobble anymore.

The authors added their "Shrinkage" tools into this balancing act to make the final stack even more stable.

5. The Test: The Simulation Lab

To prove their method works, the authors didn't just guess; they ran thousands of computer simulations (Monte Carlo simulations).

They created fake data with known "hidden groups" and "random quirks."
They introduced "collinearity," which is like having two variables that are almost identical (e.g., measuring height in inches and centimeters). This usually confuses computers.
They tested their three tools (Ridge, Lasso, Elastic Net) against this messy data.

The Results:
The paper claims that Lasso and Elastic Net generally performed better than Ridge. They were more accurate at finding the true hidden groups, especially when the data was very messy or when the variables were highly correlated. In the "Baseball Salary" example, the Lasso and Elastic Net methods gave more reliable predictions for player salaries based on their performance stats.

Summary

The paper is essentially saying: "When you have a complex crowd made of different subgroups with individual quirks, don't use a simple average. Use our new 'Shrinkage' method to gently pull your estimates toward the truth. And if you have to choose a tool, the Elastic Net or Lasso seems to be the most reliable way to sort out the noise and find the real patterns."

They tested this on simulated data and real-world examples like baseball player salaries and schoolgirls' heights, showing that their method helps separate the signal from the noise more effectively than older methods.

Technical Summary: Shrinkage Estimators in Finite Mixture of Mixed Generalized Estimating Equations

Problem Statement
The paper addresses the challenge of modeling data arising from heterogeneous populations where individuals within subgroups exhibit correlation due to repeated measurements or random effects. Standard linear mixed models (LMM) assume a single homogeneous population structure, which is inappropriate when the underlying population consists of distinct subgroups with different regression parameters. While Finite Mixture Models (FMM) handle population heterogeneity and Linear Mixed Models handle within-subject correlation, existing literature often treats these separately. The authors identify a gap in estimating parameters for a Finite Mixture-Mixed Model (FMMM), which combines both features: a mixture of linear mixed models. Furthermore, in high-dimensional settings or when multicollinearity exists among predictors, standard estimation methods (like Maximum Likelihood or standard Least Squares) may yield unstable or inefficient estimates. The paper proposes a solution that integrates shrinkage estimation techniques into the FMMM framework to improve parameter stability and selection.

Methodology
The authors propose a novel estimation framework based on the Least Square Principle combined with Shrinkage Methods. The core methodology involves the following steps:

Model Definition: The FMMM is defined where the response variable $y_{ti}$ for the $i$ -th observation at factor level $t$ is a mixture of $K$ linear mixed components. Each component $j$ has fixed effects $\beta_j$ , random effects $\alpha_{tj}$ , and mixing probabilities $\pi_j$ . The model assumes normal distributions for random effects and errors.
Iterative Weighted Least Squares (IWLS): Building on the work of Rezazadeh et al. (2020), the authors adapt the IWLS algorithm for FMMM. Since the variance of errors is not fixed and depends on the mixture structure, the model is transformed using a weighting matrix derived from the covariance structure ( $\Sigma^{-1/2}$ ). The algorithm iteratively estimates the mixing probabilities ( $\pi$ ) and the fixed effect vectors ( $\beta$ ) by solving normal equations derived from the weighted least squares objective function.
Shrinkage Integration: To address potential multicollinearity and overfitting, the authors incorporate three specific shrinkage estimators into the IWLS objective function:
- Ridge Regression: Adds an $L_2$ penalty ( $\sum \beta^2$ ) to the least squares criterion.
- Lasso Regression: Adds an $L_1$ penalty ( $\sum |\beta|$ ) to induce sparsity.
- Elastic Net: Combines both $L_1$ and $L_2$ penalties.
  The objective function is modified to include these penalty terms, weighted by the mixing probabilities.
Tuning Parameter Selection: The regularization parameters for the shrinkage methods are selected using the Cross-Validation (CV) method.
Evaluation Metric: The performance of the estimators is evaluated using the Generalized Mean Square Error (GMSE), calculated via Monte Carlo simulations.

Key Contributions

Model Formulation: The paper formally defines the Finite Mixture-Mixed Model (FMMM) and derives its matrix notation, expectation, and variance-covariance structures.
Algorithmic Extension: It extends the IWLS algorithm, previously used for standard mixture models, to the more complex FMMM context by incorporating shrinkage penalties directly into the iterative estimation of fixed effects.
Comparative Analysis: The study provides a systematic comparison of Ridge, Lasso, and Elastic Net estimators within the FMMM framework, a combination not extensively covered in prior literature for this specific model type.

Results
The authors conducted extensive Monte Carlo simulations and analyzed two real-world datasets to evaluate the proposed methods:

Simulation Studies:
- Simulations were run with varying sample sizes ( $n=50, 100, 200$ ), predictor dimensions ( $p=8, 12$ ), and correlation levels ( $\rho = 0.75, 0.90, 0.99$ ).
- Performance: The results indicate that Lasso and Elastic Net generally outperform Ridge regression in terms of GMSE, particularly as the degree of multicollinearity ( $\rho$ ) increases.
- Component Sensitivity: The performance varied slightly based on the number of mixture components ( $K$ ). For $K=3$ , Elastic Net showed superior performance, while for $K=2, 4, 5, 6$ , Lasso often performed better than Elastic Net and Ridge.
- High Collinearity: In scenarios with very high correlation ( $\rho=0.99$ ), the Ridge estimator's GMSE increased significantly, whereas Lasso and Elastic Net remained robust.
Real Data Applications:
- Baseball Salaries: Analyzing 1992 salaries of 337 players, the authors used log(salary) as the response. Lasso and Elastic Net yielded lower GMSE (16.955 and 17.035, respectively) compared to Ridge (25.471).
- Schoolgirls Heights: Analyzing growth data of 20 girls, Lasso and Elastic Net again demonstrated lower GMSE (19.777 and 19.896) compared to Ridge (21.349).

Significance and Claims
The paper claims that combining the IWLS approach with shrinkage methods provides a robust framework for estimating parameters in Finite Mixture-Mixed Models. The authors assert that their proposed iterative shrinkage approach effectively handles the heterogeneity of populations and the correlation within individuals simultaneously.

The primary significance lies in the empirical finding that shrinkage methods (specifically Lasso and Elastic Net) offer superior accuracy and stability over standard Ridge regression when estimating FMMM parameters, especially in the presence of multicollinearity. The study positions itself as a continuation of Rezazadeh et al. (2020), extending their IWLS work to include regularization, thereby offering a practical solution for complex data structures common in social sciences, medicine, and economics where both population subgroups and random effects are present. The authors modestly conclude that their simulation results support the adoption of these shrinkage estimators for improved parameter estimation in FMMM contexts.