📊 statistics

Finite mixture representations of zero-and- $N$ -inflated distributions for count-compositional data

This paper introduces a unifying finite mixture framework for two multivariate models handling zero-inflation in count-compositional data, deriving their statistical properties and developing enhanced Bayesian inference schemes validated through simulations and a gut microbiome application.

Original authors: André F. B. Menezes, Andrew C. Parnell, Keefe Murphy

Published 2026-03-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: André F. B. Menezes, Andrew C. Parnell, Keefe Murphy

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 understand a very specific type of puzzle: Count-Compositional Data.

The Puzzle: The "Fixed Total" Problem

Think of a pizza cut into $N$ slices. You have a group of friends (categories) who want to eat these slices.

The Rule: The total number of slices eaten must always equal $N$ . If you eat 3 slices, your friend can only eat $N-3$ .
The Problem: In real life, some friends often eat zero slices. Sometimes, they eat zero because they aren't hungry (a "sampling zero"). But sometimes, they eat zero because they simply don't exist in that group (a "structural zero").
The Complication: If a friend eats zero slices, those slices don't disappear; they get redistributed to the others. If everyone except one friend eats zero, that one lucky friend gets the entire pizza ( $N$ slices).

Standard math models (like the Multinomial distribution) assume everyone has a fair chance of eating a slice. But when you have too many zeros (or too many people eating the whole pizza), the standard models break down. They can't explain why the data looks so "spiky" at zero or at the maximum number.

The Solution: Two New "Mixing" Recipes

The authors of this paper invented two new statistical recipes to solve this. They call them ZANIM and ZANIDM.

Think of these recipes not as single dishes, but as smoothies made by mixing different ingredients.

1. The "Standard" Smoothie (The Multinomial Part)

This is the normal scenario where everyone eats a few slices based on their usual appetite.

2. The "Zero" Smoothie (The Inflation Part)

This is a special ingredient added to the mix. It represents the "structural zeros"—the friends who are definitely not eating.

The Magic: The authors realized that instead of just saying "there are too many zeros," you can mathematically treat the data as a mixture.
The Analogy: Imagine you have a bag of marbles.
- Sometimes, you pull out a "Zero Marble" (the friend eats nothing).
- Sometimes, you pull out a "Full Pizza Marble" (the friend eats everything because everyone else ate nothing).
- Most of the time, you pull out a "Normal Marble" (everyone shares the pizza).

The paper proves that these complex, messy data sets can be perfectly described as a finite mixture of these simple scenarios.

The Two New Recipes

Recipe A: ZANIM (The Simple Mix)

What it is: A mixture of standard pizza-sharing scenarios and "Zero/Full" scenarios.
Best for: When the variation in how people eat is mostly due to the "Zero/Full" rules. It's like a smoothie made of water and fruit juice. Simple and effective.

Recipe B: ZANIDM (The Spicy Mix)

What it is: This is the same as ZANIM, but it adds a "spicy" ingredient called Overdispersion.
The Metaphor: Imagine the pizza slices aren't just fixed; they are slightly wobbly. Sometimes the appetite of the group varies wildly. One day everyone is starving; the next day, no one is hungry.
Why it matters: Real-world data (like bacteria in a gut) is often "wobbly" (overdispersed). ZANIDM handles this extra chaos better than ZANIM. It's like adding a complex spice blend to your smoothie to handle different tastes.

Why This Matters: The "Gut Feeling"

The authors tested these recipes on a real-world dataset: Human Gut Microbiome.

The Data: They looked at 28 different types of bacteria in 98 people.
The Issue: Many people had zero of certain bacteria. Was it because the bacteria were dead (structural zero)? Or just because the test didn't catch them (sampling zero)? Or was the bacteria just very rare and variable (overdispersion)?
The Result: The new "ZANIDM" recipe (the spicy mix) fit the data much better. It correctly identified that for some bacteria, the zeros were real (structural), while for others, the zeros were just due to the wild variability of the gut environment.

The "Secret Sauce": Better Math for Computers

The paper also introduced a new way for computers to "learn" these recipes (Bayesian Inference).

Old Way: The computer had to guess and check, jumping back and forth between different possibilities, which was slow and clumsy (like trying to find a needle in a haystack by looking at one straw at a time).
New Way: The authors developed a "collapsed" method. It's like using a magnet to pull all the needles out at once. It's faster, more accurate, and requires less computing power.

Summary

In short, this paper says:

Count data with lots of zeros is tricky because zeros change the rules for everyone else.
We can fix this by viewing the data as a mixture of simple scenarios (Zero, Full, and Normal).
We have two tools: A simple one (ZANIM) and a flexible one for messy data (ZANIDM).
We have a faster way to use these tools on computers.
It works in real life, helping scientists understand the complex world of bacteria in our guts.

It's essentially a new, smarter way to count things when the "nothing" and "everything" scenarios are just as important as the "some" scenarios.

1. Problem Statement

Count-compositional data consists of vectors of non-negative integers $(y_1, \dots, y_d)$ summing to a fixed total $N$ (trials). Standard models like the Multinomial (MN) and Dirichlet-Multinomial (DM) distributions often fail to capture excess zeros (zero-inflation) common in fields like microbiome sequencing.

The Challenge: In multivariate settings, zero-inflation is complex. If one category is zero, the remaining categories must sum to $N$ . If multiple categories are zero, the remaining non-zero categories effectively "inflate" to compensate.
The Gap: Existing literature often addresses zero-inflation by modifying the latent space (e.g., zero-inflated probabilities) without fully characterizing the resulting sampling distribution of the counts. Furthermore, previous models often fail to explicitly account for N-inflation (where a single category takes the value $N$ because all others are zero) or treat it as a separate phenomenon.
Goal: To develop a unified probabilistic framework that characterizes both zero-inflation and N-inflation as a finite mixture of standard distributions, enabling the derivation of theoretical properties and efficient Bayesian inference.

2. Methodology

The authors propose two new distributions, ZANIM (Zero-and-N-inflated Multinomial) and ZANIDM (Zero-and-N-inflated Dirichlet-Multinomial), derived via a unifying finite mixture framework.

A. Derivation of Distributions

The core methodology involves expressing the likelihood as a product of Poisson-type terms (using an auxiliary Gamma variable $\phi$ ) and introducing a zero-inflation mechanism.

ZANIM (Multinomial-based):
- Starts with the Multinomial likelihood augmented by a Gamma variable.
- Introduces a zero-inflation parameter $\zeta_j$ for each category, creating a "ZI-Poisson-type" form.
- Result: The marginal distribution of counts is a finite mixture of $2^d$ components. These components include:
  - A standard Multinomial distribution (no zeros).
  - Reduced-dimension Multinomials (where specific subsets of categories are zero).
  - Degenerate distributions at $0$ (all categories zero) and at $N$ (one category is $N$ , others are $0$).
- This provides a proper sampling distribution for the counts, not just a latent representation.
ZANIDM (Dirichlet-Multinomial-based):
- Extends the ZANIM framework by compounding the Multinomial with a Dirichlet distribution (via Gamma latent variables $\lambda_j$ ).
- Result: A finite mixture of $2^d$ components where the non-degenerate parts are Dirichlet-Multinomial (DM) distributions (or reduced-dimension DMs).
- This formulation fully characterizes the ZANIDM (previously known as ZIDM) as a sampling distribution, explicitly modeling overdispersion alongside zero/N-inflation.

B. Theoretical Properties

Using the finite mixture representation, the authors derive:

Marginal Distributions: The marginal distribution of any single category $Y_j$ is a mixture of degenerate masses (at 0 and $N$ ) and Binomial (for ZANIM) or Beta-Binomial (for ZANIDM) distributions.
Moments: Closed-form expressions for means, variances, and covariances are derived.
- Crucially, unlike standard MN/DM models which enforce negative covariance, the mixture structure allows for positive covariance between categories, a feature observed in real-world data.
Stochastic Representations: Efficient algorithms for generating random variates are provided based on Bernoulli indicators determining which mixture component is active.

C. Bayesian Inference

The paper develops efficient Bayesian inference schemes using data augmentation and Gibbs sampling.

ZANIM: Uses latent variables for zero-inflation indicators ( $z$ ) and the auxiliary Gamma variable ( $\phi$ ). This leads to conjugate updates for parameters $\theta$ and $\zeta$ .
ZANIDM: Improves upon existing methods (e.g., Koslovsky [10]) by proposing a collapsed Gibbs sampling approach.
- Instead of joint "expand-and-contract" Metropolis-Hastings steps for latent variables, the authors marginalize out the latent Gamma variables ( $\lambda$ ) to update the zero-indicators ( $z$ ) directly.
- This allows for fast conjugate updates for $\zeta$ and $\alpha$ (concentration parameters), significantly improving MCMC efficiency.
- Three methods for sampling $\alpha$ are evaluated: Data Augmentation with Power-Truncated Normal (DA-PTN), Random Walk Metropolis-Hastings (MH-RW), and Slice Sampling (SS).

3. Key Contributions

Unified Framework: Established a unifying finite mixture framework that represents both ZANIM and ZANIDM as sampling distributions with $2^d$ components, clarifying the relationship between zero-inflation and N-inflation.
Novel Distribution (ZANIM): Introduced the ZANIM distribution, which had not been previously characterized as a sampling distribution of counts.
Re-characterization of ZANIDM: Provided a complete probabilistic characterization of the ZANIDM (ZIDM) distribution, moving beyond latent stochastic representations to explicit PMFs.
Efficient Inference: Developed collapsed Gibbs sampling algorithms that outperform existing MCMC methods (specifically those by Koslovsky) in terms of Effective Sample Size (ESS) and computational speed.
Theoretical Insights: Derived moments and marginal distributions, demonstrating that these models can capture positive correlations and handle overdispersion better than standard models.

4. Results

Simulation Studies

MCMC Efficiency: The proposed collapsed Gibbs sampling for ZANIDM consistently yielded higher ESS ratios compared to the standard ZIDM algorithm, particularly for the concentration parameters ( $\alpha$ ).
Model Recovery:
- When data was generated from ZANIM, both models recovered parameters well, though ZANIDM showed wider credible intervals for $\alpha$ (due to model mismatch).
- When data was generated from ZANIDM, the ZANIDM model provided superior fit (higher ELPD) and better parameter recovery. The ZANIM model struggled to capture the overdispersion inherent in the DM component, leading to biased estimates of zero-inflation parameters.
Model Selection: Expected Log-Predictive Density (ELPD) analysis confirmed that the correct distribution (ZANIM vs. ZANIDM) should be chosen based on the data's overdispersion characteristics.

Real Data Application (Human Gut Microbiome)

Dataset: 98 individuals, 28 genera-level taxa, with 32.6% zero counts and high overdispersion.
Findings:
- Both models converged well.
- Interpretation of Zeros: For taxa with high overdispersion (e.g., Actinomycineae), the ZANIDM model attributed zeros largely to overdispersion (low $\zeta$ , low $\alpha$ ), whereas ZANIM attributed them to structural zero-inflation (high $\zeta$ ).
- Fit: The ZANIDM model provided a significantly better fit (higher ELPD) to the microbiome data, suggesting that the overdispersion in this dataset is driven by the Dirichlet-multinomial nature of the data rather than purely structural zeros.
- Trace Plots: Demonstrated stable mixing and convergence even with highly imbalanced trial numbers ( $N$ ) and zero-inflation levels.

5. Significance

Theoretical Advancement: The paper bridges the gap between latent variable models and explicit sampling distributions for count-compositional data, providing a rigorous mathematical foundation for zero-and-N-inflated models.
Practical Utility: The proposed inference schemes are computationally efficient and robust, making these complex mixture models accessible for high-dimensional data (e.g., microbiome, ecology).
Flexibility: The framework allows for the modeling of positive correlations between categories, a critical feature often missed by standard multinomial models.
Future Directions: The authors suggest extending the framework to include covariates (regression) and other component distributions (e.g., Conway-Maxwell-Multinomial), as well as relaxing the independence assumption of the zero-inflation indicators.

In summary, this work provides a comprehensive statistical toolkit for analyzing zero-inflated count-compositional data, offering both theoretical clarity and practical computational advantages over existing methods.