📊 statistics

Addressing outliers in mixed-effects logistic regression: a more robust modeling approach

This study proposes and validates a Bayesian outlier-robust mixed-effects logistic regression model using a t-distributed latent variable to effectively analyze hierarchically structured bounded count data, demonstrating superior performance and reliability over conventional approaches in the presence of outliers through both simulation and a longitudinal medication adherence case study.

Original authors: Divan A. Burger, Sean van der Merwe, Emmanuel Lesaffre

Published 2026-02-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Divan A. Burger, Sean van der Merwe, Emmanuel Lesaffre

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 trying to predict how many days a person will take their heart medication in a month. You have data from hundreds of patients, and you want to find the "average" behavior to understand if a new counseling program helps.

Usually, statisticians use a standard ruler (a Binomial model) to measure this. But real life is messy. Sometimes a patient forgets a whole week because they went on vacation, or they take every single pill because they are feeling anxious. These "weird" days are called outliers.

If you use the standard ruler, these weird days can skew your results, making the average look wrong. It's like trying to measure the average height of a basketball team, but one person is a 7-foot giant and another is a 3-foot child; the average might look like a 5-foot-6 person, which doesn't represent the team well.

This paper introduces a new, smarter ruler called the Binomial-Logit-T model. Here is how it works, explained simply:

1. The Problem: The "Perfect" Ruler Breaks

Most models assume that if a patient is supposed to take a pill, they either do it or they don't, and the variation is predictable. But human behavior is unpredictable.

The Old Way: If a patient has a few "bad days" (outliers), the standard model gets confused. It tries to stretch the whole curve to fit those bad days, which distorts the picture for everyone else.
The Result: You might think the medication program is failing when it's actually working, or vice versa.

2. The Solution: The "Stretchy, Shock-Absorbing" Ruler

The authors built a model that uses a t-distribution. Think of this as a shock absorber on a car.

Standard Model (Binomial): Like a rigid steel beam. If you hit a bump (an outlier), the whole car shakes violently.
New Model (Binomial-Logit-T): Like a car with heavy-duty suspension. When it hits a bump (an outlier), the suspension absorbs the shock. The car keeps driving smoothly, and the other passengers (the rest of the data) aren't thrown around.

This "shock absorber" allows the model to say, "Okay, this patient had a crazy week, but we know that's an exception. Let's not let it ruin our calculation for the whole group."

3. The Secret Weapon: The "Pseudo-Median"

In statistics, there are two ways to find the "center" of a group: the Mean (average) and the Median (the middle person).

The Mean is easily tricked by outliers (like the 7-foot giant in the basketball team).
The Median is tough; it ignores the extremes.

The authors' model is special because it calculates a "Pseudo-Median."

The Analogy: Imagine you are trying to guess the middle number in a list of lottery tickets. Usually, you have to do a massive, complicated math puzzle to find the true middle.
The Innovation: This new model gives you a "shortcut" (a closed-form formula) that gets you so close to the true middle that the difference is less than one single pill. It's like having a GPS that tells you you are "within one block" of your destination without needing to drive there first. This makes the results easy to explain to doctors and patients.

4. The Real-World Test: The Medication Study

The authors tested this on real data from 392 patients taking cholesterol medication.

The Data: Some patients were perfect; some were chaotic.
The Result: The new model (the shock absorber) handled the chaotic patients much better than the old models. It gave a clearer, more honest picture of how the medication program was working.
The Proof: They ran computer simulations where they intentionally added "fake" bad data. The new model stayed calm and accurate, while the old models went haywire.

5. Why This Matters

In the real world, data is rarely perfect. People forget, get sick, or have emergencies.

Old Approach: "Let's throw out the weird data points so the math works." (This is dangerous because you might throw away important information).
New Approach: "Let's keep all the data, but use a model that is strong enough to handle the weirdness."

In a nutshell: This paper gives statisticians a super-strong, flexible tool to analyze messy, real-world data without getting thrown off by the outliers. It ensures that when we make decisions about healthcare or policy, we are looking at the true picture, not a distorted one caused by a few bad days.

1. Problem Statement

The paper addresses the challenge of analyzing hierarchically structured bounded count data (e.g., medication adherence counts out of a total number of days) that exhibit two specific statistical issues:

Overdispersion: The observed variance exceeds the mean, violating the assumptions of standard binomial models.
Outliers: Data points with abrupt deviations or extreme values (often caused by behavioral inconsistencies in longitudinal studies) that distort parameter estimates in traditional models.

Existing solutions have limitations:

Standard Binomial: Fails to handle overdispersion or outliers.
Beta-Binomial & Binomial-Logit-Normal: Handle overdispersion but lack robustness against outliers, as they rely on light-tailed distributions (Beta or Normal) for the latent variable.
Zero-Inflated Models: Address extreme zeros/ones but fail to handle outliers that fall within the intermediate range of the data.

The authors argue that current mixed-effects models for bounded counts do not adequately balance robustness to outliers with the interpretability of central tendency measures (mean vs. median).

2. Methodology

The authors propose a Binomial-Logit-T Mixed-Effects Model implemented within a Bayesian framework using JAGS (Just Another Gibbs Sampler).

Core Model Structure

The model defines the response variable $Y_{ij}$ (bounded count) as:
$Y_{ij} \sim \text{Binomial}(m_{ij}, p_{ij})$
where the probability of success $p_{ij}$ is linked to a latent variable $\xi_{ij}$ via the logistic function:
$p_{ij} = \frac{1}{1 + e^{-\xi_{ij}}}$

The Innovation: Unlike the Binomial-Logit-Normal model where $\xi_{ij}$ follows a Normal distribution, the proposed model assumes $\xi_{ij}$ follows a Student's t-distribution:
$\xi_{ij} \sim t(\eta_{ij}, \sigma^2, \nu)$

$\eta_{ij}$ : The linear predictor ( $\mathbf{x}_{ij}'\boldsymbol{\beta} + \mathbf{z}_{ij}'\mathbf{u}_i$ ), containing fixed effects ( $\boldsymbol{\beta}$ ) and cluster-specific random effects ( $\mathbf{u}_i$ ).
$\sigma^2$ : Scale parameter controlling dispersion.
$\nu$ : Degrees of freedom. Low $\nu$ values create heavy tails, providing robustness against outliers.

Key Technical Features

Observation-Level Random Effects (OLRE): The latent variable $\xi_{ij}$ acts as an observation-level random effect, allowing the model to capture extra-binomial variation and outlier behavior simultaneously.
Pseudo-Median: A critical theoretical contribution is the derivation of a closed-form pseudo-median for the response:
$\tilde{M}_i = m_i \cdot \text{logit}^{-1}(\eta)$
The authors prove that this quantity differs from the true discrete median by strictly less than one count. This allows for direct interpretation of fixed effects without numerical integration, unlike the mean.
Mean Estimation: Calculating the marginal mean requires integrating out the latent t-distributed variable, necessitating numerical integration or simulation-based approaches.

Comparison Models

The proposed model is compared against three competitors:

Binomial: No overdispersion, no outlier robustness.
Beta-Binomial: Handles overdispersion, no outlier robustness.
Binomial-Logit-Normal: Handles overdispersion, no outlier robustness.

3. Key Contributions

Novel Distribution: Introduction of the Binomial-Logit-T distribution for mixed-effects bounded count data, specifically designed to handle heavy tails and outliers.
Closed-Form Interpretability: Derivation of a pseudo-median that is a closed-form function of the linear predictor. This offers a robust measure of central tendency that is easier to interpret than the mean (which requires integration) and more accurate than the true discrete median in the presence of skewness.
Bayesian Implementation: A fully Bayesian implementation using JAGS that avoids the need for deriving asymptotic properties of estimators, facilitating complex model specification.
Gap Filling: Addresses the lack of literature on outlier-robust mixed models specifically for bounded count data (distinct from zero-inflated models).

4. Results

The study validates the model using a longitudinal medication adherence dataset (392 patients) and a simulation study.

Real-World Application (Medication Adherence)

Model Fit: The Binomial-Logit-T model achieved the lowest WAIC (Watanabe-Akaike Information Criterion), indicating superior expected out-of-sample performance compared to Beta-Binomial, Binomial-Logit-Normal, and standard Binomial models.
Outlier Detection: Kullback-Leibler (K-L) divergence analysis showed the Binomial-Logit-T model had the lowest divergence values, confirming its resilience to influential observations. In contrast, the standard Binomial model showed high sensitivity to outliers.
Residual Diagnostics: Using the DHARMa package, the Binomial-Logit-T model passed uniformity, dispersion, and outlier tests (p-values > 0.05), whereas all other models failed the outlier and/or uniformity tests.
Parameter Estimates: The estimated degrees of freedom ( $\nu$ ) for the t-distribution was less than 2, confirming the presence of heavy tails/outliers in the adherence data.
Uncertainty: The non-robust models (Beta-Binomial, Logit-Normal) produced wider 95% Highest Posterior Density (HPD) intervals for adherence proportions, indicating greater uncertainty due to their struggle with outliers.

Simulation Study

The Binomial-Logit-T model demonstrated lower Root Mean Square Error (RMSE) and shorter average 95% HPD interval lengths compared to the Binomial-Logit-Normal model when data was contaminated with outliers.
The model maintained appropriate bias and coverage properties, confirming good frequentist properties under the Bayesian framework.

5. Significance and Implications

Robustness without Data Removal: The approach allows researchers to retain the integrity of the dataset by modeling outliers rather than discarding them, preventing biased conclusions in fields like epidemiology and behavioral science.
Interpretability: The closed-form pseudo-median provides a practical tool for reporting results in skewed, overdispersed data where the mean is sensitive to extremes.
Computational Efficiency: Despite the complexity of the t-distribution, the model converges efficiently (approx. 20 minutes on a standard laptop), making it a viable alternative to standard mixed models.
Future Directions: The authors suggest extending the random effects structure to a multivariate t-distribution to handle cluster-level outliers and exploring skew-t distributions for directional outliers, though this would sacrifice the closed-form pseudo-median.

In conclusion, the paper presents a statistically rigorous and practically effective framework for analyzing bounded count data in the presence of outliers, offering a significant improvement over traditional generalized linear mixed models.