📊 statistics

A Compound Logistic Regression Model for Binary Responses

This paper introduces the compound logistic regression model, which extends traditional logistic regression by allowing for correlated responses and covariates through a mean response function composed of multiple logistic components, thereby overcoming the limitation of fixed 0 and 1 asymptotes.

Original authors: Anthony Almudevar, Jacob Almudevar

Published 2026-02-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anthony Almudevar, Jacob Almudevar

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 whether a light switch will turn a lamp on or off. In the world of statistics, this is called a "binary response" (On/Off, Yes/No, 1/0).

For decades, the standard tool for this job has been Logistic Regression. You can think of this standard tool as a very smooth, S-shaped ramp. As you push a lever (a variable like a biomarker), the probability of the light turning on slowly rises from 0% to 100%.

The Problem:
The authors of this paper point out a flaw in this standard ramp. In the real world, things don't always go from 0% to 100%.

Sometimes, even if you push the lever all the way, the light only gets to 80% bright (it never fully turns on).
Sometimes, even if you don't push the lever at all, the light is already flickering at 10% (it never fully turns off).

The standard "S-ramp" is stuck with its ends fixed at 0 and 1. It's too rigid for many real-life situations, like predicting infection rates or disease risks where a "floor" and "ceiling" exist that aren't zero or one.

The Solution: The Compound Logistic Model
The authors, Anthony and Jacob Almudevar, propose a new, more flexible machine called the Compound Logistic Regression Model.

Instead of one single S-shaped ramp, imagine building a machine out of three smaller, separate S-shaped ramps that work together to create the final result.

Here is how they combine them, using a Vaccine Analogy from the paper:

Ramp A (The Exposure): This ramp predicts the chance that a person gets exposed to a virus.
Ramp B (The Protection): This ramp predicts how well a vaccine works (efficacy).
Ramp C (The Threshold): This ramp predicts the level of antibodies needed to trigger that protection.

In the old model, you had to guess how these interacted in a messy way. In the new Compound Model, the authors create a "recipe" (a mathematical function called $H$ ) that mixes these three ramps together.

If you have high exposure (Ramp A) but low protection (Ramp B), the final risk is high.
If you have high exposure but high protection, the final risk drops.

Why is this "Compound" approach better?
The paper argues that this method is like having a team of specialists rather than one generalist.

Specialization: You can have one set of data (covariates) that only affects the "Exposure" ramp, and a completely different set of data that only affects the "Protection" ramp.
Flexibility: You can change the "floor" and "ceiling" of your prediction. For example, in a study of diabetes, the model showed that even with high blood sugar, people with asthma had a different "ceiling" (maximum risk) than those without asthma. The old model couldn't easily show this; the new model handles it naturally.

The "Correlated" Twist
The paper also handles a tricky situation: what if the data points aren't independent?
Imagine you are measuring the health of a family. The parents and children share genes and environment, so their results are "correlated" (related to each other). Standard models often treat every person as a stranger. This new model includes a way to account for these family-like connections, ensuring the math doesn't get confused by the relationships between the data points.

The "Stability" Trick (Regularization)
The authors found that when they tried to fit this complex machine to real data, the math sometimes got "wobbly" and failed to find a solution (it wouldn't converge).
To fix this, they added a "shock absorber" called Regularization. Think of this as a gentle hand that keeps the machine from wobbling too wildly. It slightly biases the answer to make the calculation stable and reliable. Their tests showed that without this shock absorber, the machine failed to work over half the time; with it, it worked every time.

Real-World Test: Diabetes and Asthma
The authors tested their new model using real data from the "MIDUS" study (a survey of American adults).

The Setup: They tried to predict Diabetes (Yes/No) based on blood sugar levels (Hemoglobin A1c).
The Twist: They also looked at whether the person had Asthma.
The Result: The model confirmed that while high blood sugar increases diabetes risk for everyone, having Asthma actually lowers the maximum risk ceiling for those with very high blood sugar. The standard model would have missed this nuance, but the "Compound" model spotted it clearly.

In Summary
This paper introduces a new statistical "Swiss Army Knife." It takes the reliable, familiar logic of logistic regression but adds extra gears and levers. This allows researchers to:

Set realistic minimums and maximums for their predictions (not just 0 and 1).
Separate different causes (like exposure vs. protection) into their own distinct parts.
Handle data where people are related to each other.
Stay stable even when the data is messy.

The authors have built a software package (an R tool called CLmodel) so other scientists can start using this new, more flexible way of looking at binary data immediately.

Technical Summary: A Compound Logistic Regression Model for Binary Responses

Problem Statement
Standard logistic regression is the dominant method for modeling binary responses, yet it possesses a significant limitation: the response function is constrained to have asymptotes fixed at 0 and 1. In many practical applications, such as modeling disease prevalence based on biomarker levels or vaccine efficacy, the probability of an event does not necessarily approach 0 or 1 even at extreme values of the predictor. Furthermore, existing flexible models that allow for variable asymptotes often struggle to incorporate correlated responses or multiple covariates in a unified, interpretable framework. The authors identify a need for a model that generalizes the logistic response to allow for arbitrary lower ( $a_L$ ) and upper ( $a_U$ ) asymptotes while retaining the analytical tractability and interpretability of standard logistic regression.

Methodology
The paper proposes the Compound Logistic Regression (CLR) model. This approach constructs a mean response function $\pi(x)$ by compounding multiple logistic regression functions through a specific compounding function $H$ .

Model Structure:
The core idea is to express the probability $\pi$ as a function of three distinct logistic components, $\phi(\eta_1), \phi(\eta_2), \phi(\eta_3)$ , where $\eta_j = \tilde{\beta}_j^T \tilde{x}_j$ .
- Asymptote Model: The authors demonstrate that the lower and upper asymptotes can be modeled as logistic functions themselves. For instance, a model where the response transitions between $a_L$ and $a_U$ based on a biomarker $z$ is formulated as:
  $\pi_c(\tilde{x}_1, \tilde{x}_2, \tilde{x}_3) = \phi(\eta_3)\phi(\eta_1) + \phi(\eta_2)(1 - \phi(\eta_1))$
  Here, $a_L = \phi(\eta_2)$ and $a_U = \phi(\eta_3)$ , allowing these limits to depend on covariates.
- Vaccine Efficacy Model: An alternative compounding function is proposed for scenarios involving exposure, protection, and efficacy, such as:
  $\pi_{inf} = \xi(1 - \gamma \kappa(z))$
  where $\kappa(z)$ models the correlate of protection (COP) threshold, $\gamma$ represents efficacy, and $\xi$ represents exposure risk. Each component is modeled via a separate logistic function.
Inference and Estimation:
- Estimating Equations: The paper utilizes a quasi-likelihood approach based on estimating equations. The gradient matrix of the compound model is shown to be a weighted sum of the gradient matrices of the individual component models (Equation 39). This linearity simplifies the derivation of the Hessian and the iterative fitting algorithm.
- Correlated Responses: The framework accommodates dependent responses (e.g., repeated measures) by defining a working correlation matrix $R_{\tilde{y}}$ . The authors specifically address exchangeable correlation, where observations within a group share a common correlation $\rho$ .
- Regularization: To address convergence issues and instability, particularly when the biomarker gradient is near zero, the authors incorporate ridge regression. A penalty term $D = \theta^T B \theta$ is added to the estimating equations, shrinking parameter estimates and stabilizing the numerical algorithm.
- Identifiability: The paper notes that for the model to be identifiable, the gradient associated with the primary biomarker must be non-zero. A reparametrization strategy is suggested to enforce a positive gradient direction if prior knowledge exists.
Implementation:
The methods are implemented in the R package CLmodel, which supports independent and exchangeable correlation structures, quasi-likelihood estimation, and regularization via a penalty parameter $\lambda$ .

Key Results

Simulation Study: A simulation study with $N=50,000$ $N = 50, 000$ replications was conducted to assess the model's performance under exchangeable correlation ( $\rho=0.1$ $ρ = 0.1$ ).
- Stability: The study highlighted that without regularization ( $\lambda=0$ ), the numerical algorithm suffered a high failure rate (55.6% non-convergence). Introducing a small penalty ( $\lambda=0.1$ ) resulted in 100% convergence.
- Bias-Variance Trade-off: While regularization introduced slight bias, it significantly reduced the variance of the standard error estimates and improved the coverage of confidence intervals, bringing them closer to the nominal 95% level.
- Parameter Recovery: The model successfully recovered true parameters, including the correlation coefficient $\rho$ , and demonstrated high power to detect covariate effects on the asymptotes.
Real Data Application: The model was applied to data from the Midlife in the United States (MIDUS) project to predict diabetes ( $y=1$ $y = 1$ ) based on Hemoglobin A1c% ( $z$ $z$ ) and asthma status ( $x$ $x$ ).
- The fitted curves showed that the probability of diabetes approaches different upper asymptotes for subjects with and without asthma.
- The parameter $\beta_{1,3}$ , governing the difference in the upper asymptote, was statistically significant ( $p=0.017$ ), indicating that subjects with asthma have a significantly lower probability of diabetes at high A1c levels (Odds Ratio = 0.11).
- The model's transition point aligned closely with American Diabetes Association guidelines (prediabetes range 5.7%–6.5%), validating the biological plausibility of the fitted curve.

Significance and Claims
The paper claims that the Compound Logistic Regression model offers a flexible extension to standard logistic regression that allows for:

Variable Asymptotes: The ability to model lower and upper bounds of the response probability that are not fixed at 0 and 1, and which can depend on covariates.
Interpretability: The model retains the intuitive interpretation of logistic regression coefficients (log-odds) for each component, allowing researchers to separately model factors affecting the threshold, the baseline risk, and the maximum risk.
Correlation Handling: The framework naturally extends to correlated data structures (e.g., longitudinal studies) using generalized estimating equations (GEE) principles.
Computational Stability: The inclusion of ridge regularization is shown to be critical for the numerical stability of the fitting algorithm in complex, multi-component scenarios.

The authors emphasize that this approach permits a "greater variety of models" while maintaining the "advantages of logistic regression," specifically citing the implementation in the CLmodel package as a practical tool for these analyses. The paper does not claim to replace standard logistic regression but rather to provide a necessary generalization for applications where the fixed asymptote assumption is inappropriate.

More like this