📊 statistics

Assessing the impact of variance heterogeneity and misspecification in mixed-effects location-scale models

This paper uses a simulation study to demonstrate that neglecting variance heterogeneity in standard Linear Mixed Models leads to biased estimates and poor coverage, while showing that in Mixed-Effects Location-Scale Models, misspecification of the scale component does not bias location estimates, though misspecifying the location component does affect scale estimates.

Original authors: Vincent Jeanselme, Marco Palma, Jessica K Barrett

Published 2026-01-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Vincent Jeanselme, Marco Palma, Jessica K Barrett

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 teacher trying to understand how students learn a new subject over time. You have a class of 300 students, and you test them every few weeks.

The Old Way (The Linear Mixed Model):
Traditionally, statisticians use a tool called a "Linear Mixed Model" (LMM). Think of this like drawing a single, straight line through the class's average test scores to see the general trend. It assumes that while every student has their own starting point (some are naturally smarter, some need more help), the amount of "noise" or unpredictability in their scores is the same for everyone.

In this old view, if Student A's scores jump around wildly and Student B's scores are very steady, the model treats them as if they are equally steady. It assumes the "wobble" in the data is constant. The paper calls this assumption homoscedasticity (a fancy word for "equal spread").

The Problem:
In real life, this assumption is often wrong. Some students might be very consistent, while others have huge ups and downs. If you ignore this difference (if you ignore heteroscedasticity), your math gets messy. You might think you are very sure about your conclusions when you actually aren't, or you might miss important patterns.

The New Tool (The MELSM):
The authors of this paper are testing a newer, more flexible tool called the Mixed-Effect Location-Scale Model (MELSM).

Location: This is the "average" line (the trend).
Scale: This is the "wobble" or variability around that line.

The MELSM doesn't just draw the average line; it also draws a separate line to predict how much each student's scores will wobble. It asks: "Does Student A's wobble depend on how much they studied? Does Student B's wobble depend on their age?"

What the Paper Found (The Experiments):
The authors ran thousands of computer simulations (like running a video game with different rules) to see what happens when you use the old tool vs. the new tool, and what happens if you make mistakes in how you set them up.

Here are the main takeaways, explained simply:

Ignoring the Wobble is Dangerous:
If you use the old tool (LMM) when the data actually has different amounts of wobble, your confidence in the results is fake. You might think you are 95% sure of an answer, but you are actually much less sure. The "error bars" on your results become too small, leading to wrong conclusions.
The "Average" vs. The "Wobble" Relationship:
- If you get the "Wobble" wrong: If you mess up the model for the variability (the scale), the estimate for the average trend (the location) stays mostly correct, but it becomes "fuzzier" (less precise). It's like trying to aim a dart at a target while the wind is blowing; you still hit the general area, but you're less sure exactly where.
- If you get the "Average" wrong: If you mess up the model for the average trend (the location), it completely ruins the estimate for the wobble. If you don't account for the main trend correctly, the leftover "noise" looks like it has a pattern that isn't really there.
Adding Extra Variables:
If you include extra variables in your model that aren't actually important (like adding "favorite color" to a math test model), it doesn't break the math. It just makes the computer work a little slower. The model is robust enough to ignore the useless stuff.
Missing the "Slopes":
Sometimes, a student's performance doesn't just start at a different point; it changes speed over time (a slope). If you forget to include this "speed change" in your model, your confidence in the results drops again, even if the average looks okay.
Real-World Test (The PBC Dataset):
The authors tested this on a real medical dataset about liver disease. They compared the old tool against the new tool.
- Result: The old tool and the new tool gave similar answers for the main factors (like age or sex).
- The Catch: The old tool showed a different, steeper curve for how the disease progressed over time compared to the new tool. This proves that ignoring the "wobble" can change how we see the story of the disease's progression.

The Bottom Line:
The paper argues that we should stop assuming everyone's data is equally "noisy." By using the MELSM, we can model both the average trend and the individual variability at the same time. This gives us a more honest picture of the data, preventing us from making confident mistakes.

In short: Don't just measure the average; measure how much the average wobbles, because ignoring the wobble can trick you into thinking you know more than you actually do.

Technical Summary: Assessing the impact of variance heterogeneity and misspecification in mixed-effects location-scale models

Problem Statement
Linear Mixed Models (LMMs) are a standard approach for analyzing longitudinal data, modeling the relationship between exposure and outcome while capturing individual variability through random effects. However, LMMs rely on the assumption of homoscedasticity—that the residual variance is constant across all observations and covariates. When this assumption is violated (heteroscedasticity), standard LMMs can produce biased estimates of random effect standard deviations and lead to incorrect inference (Type I and II errors) due to underestimated standard errors. While Mixed-Effect Location-Scale Models (MELSMs) offer a framework to model both the location (mean) and scale (residual standard deviation) as functions of covariates and random effects, the consequences of misspecifying these models are not fully understood. Specifically, the interplay between misspecification in the location submodel versus the scale submodel, and how such errors propagate to fixed and random effect estimates, requires further investigation.

Methodology
The authors conducted a comprehensive simulation study using synthetic longitudinal data generated based on covariates from the Primary Biliary Cholangitis (PBC) dataset. The data generation process involved:

Population: A synthetic population of $N$ individuals with baseline covariates (age, albumin, triglycerides, platelets) drawn from a multivariate normal distribution.
Longitudinal Structure: Repeated measures per individual with varying numbers of encounters ( $M$ ), where age was the primary time-varying covariate.
True Model: A correctly specified MELSM where both the location ( $y$ ) and scale ( $\omega$ ) were modeled as functions of specific covariates and random intercepts, with a correlation structure allowed between random effects.

The study evaluated the robustness of MELSMs and LMMs under various misspecification scenarios, including:

Ignoring Heteroscedasticity: Comparing a standard LMM (constant scale) against a correctly specified MELSM.
Covariate Misspecification: Adding irrelevant covariates, omitting relevant covariates, or omitting random effects in either the location or scale submodels.
Functional Form Misspecification: Fitting a linear trend when the true data generation process included non-linearity (sinusoidal transformation).
Random Effect Structure Misspecification: Neglecting random slopes when they were present in the data generation process.
Distributional Misspecification: Assuming Gaussian random effects when the true distribution was Student's $t$ .

Estimates were evaluated across 500 repetitions using Bayesian inference (via the brms package in R), focusing on bias, variance, and 95% credible interval coverage for fixed effects and random effect standard deviations. A real-data case study on the PBC dataset was also performed to illustrate practical implications.

Key Results
The simulation study yielded several critical findings regarding model misspecification:

Impact of Ignoring Heteroscedasticity: Using a standard LMM when heteroscedasticity is present leads to a significant loss of coverage for location coefficients (standard errors are underestimated) and biases the estimates of random effect standard deviations. While the point estimates for fixed effects in the location model remained centered on the truth, their precision was compromised.
Asymmetry in Misspecification (Location vs. Scale):
- Scale Misspecification: Errors in the scale model (e.g., omitting covariates or random effects) generally do not bias the fixed effect estimates of the location model. However, they increase the variance of these estimates and can bias the estimated standard deviations of the location random effects.
- Location Misspecification: Errors in the location model (e.g., omitting a relevant covariate or using a linear form for a non-linear trend) directly bias the fixed effect and random effect estimates of the scale model. This occurs because unexplained variation in the location model is absorbed into the residual variance, distorting the scale estimates.
Non-linearity: Misspecifying a non-linear time trend as linear in the location model biases the location fixed effects but has minimal impact on the scale submodel if the scale is correctly specified.
Random Effects: Neglecting a relevant random slope (when one exists in the data) primarily affects the coverage of the corresponding fixed slope estimate, often leading to under-coverage, but does not necessarily bias the fixed effect point estimate.
Distributional Assumptions: Assuming a Gaussian distribution for random effects when the true distribution is Student's $t$ did not significantly impact the estimation of fixed effects or the standard deviation of the random effects, as the Gaussian model adjusted its scale to match the variance of the true distribution.
Over-specification: Including all available covariates (even irrelevant ones) did not invalidate inference, provided the true covariates were included, though it increased computational time.

Significance and Claims
The paper concludes that modeling heteroscedasticity is crucial for valid inference in longitudinal studies. The authors assert that MELSMs provide a robust framework for handling variance heterogeneity, but the validity of conclusions depends heavily on the correct specification of the location model.

Key claims include:

Interdependence: There is a critical interplay between the location and scale models; misspecification in the location model propagates bias to the scale estimates, whereas scale misspecification primarily affects the precision (variance) of location estimates.
Practical Necessity: Even with imperfect knowledge of the scale structure, modeling residual variance is superior to assuming homoscedasticity, as the latter leads to underestimated standard errors and incorrect coverage probabilities.
Broader Implications: The findings suggest that the importance of modeling heteroscedasticity extends beyond standard LMMs to Generalized Linear Mixed Models (GLMMs) for non-normal outcomes and joint models with survival data, where within-individual variability is often a key parameter of interest.

The authors note limitations, including the reliance on specific parameter settings, the assumption of uninformative visit processes, and the focus on distributional shape rather than the prediction of individual random effects. They suggest future work should explore joint models for within-individual variability and informative observation processes.

More like this