Technical Summary: A Contaminated Model for Overdispersed Multinomial Microbiome Count Data

1. Problem Statement

Multinomial count data, particularly in the context of human gut microbiome sequencing, frequently exhibit overdispersion where observed variability exceeds that predicted by the standard multinomial distribution. The Dirichlet-multinomial (DM) distribution is the canonical solution for this overdispersion; however, it remains sensitive to anomalous observations (contamination).

In real-world datasets, such anomalies can arise from experimental noise or biological outliers. When present, these observations disproportionately inflate the estimated dispersion parameter of the DM model. This distortion leads to biased parameter estimates and potentially misleading biological inferences. While "gross anomalies" (unpredictable observations) are typically removed, "mild anomalies" (observations from a population differing from the assumed model) require a modeling approach that accommodates them without discarding data. Existing DM models lack the flexibility to distinguish between genuine overdispersion and contamination-induced variance.

2. Methodology

2.1. The Contaminated Dirichlet-Multinomial (CDM) Distribution

The authors propose the Contaminated Dirichlet-Multinomial (CDM) distribution, a two-component mixture model designed to handle mild anomalies. The probability mass function (PMF) is defined as:

$$f_{CDM_m}(y; \pi, \sigma, \delta, \eta) = (1 - \delta) f_{DM_m}(y; \pi, \sigma) + \delta f_{DM_m}(y; \pi, \eta\sigma)$$

Where:

• Reference Component: $(1 - \delta) f_{DM_m}(y; \pi, \sigma)$ represents the typical data, following a standard DM distribution with mean proportions $\pi$ and dispersion $\sigma$.
• Contaminant Component: $\delta f_{DM_m}(y; \pi, \eta\sigma)$ represents the anomalous data. It shares the same mean vector $\pi$ but possesses an inflated dispersion parameter $\eta\sigma$, where $\eta > 1$.
• Parameters:
  • $\pi$: Vector of mean proportions (interpretable as population allele frequencies or relative abundances).
  • $\sigma$: Dispersion parameter for the reference component.
  • $\delta \in (0, 1)$: The proportion of observations belonging to the contaminant component.
  • $\eta > 1$: The inflation factor representing the degree of increased variability in the contaminant component.

2.2. Identifiability

The paper addresses the identifiability of the CDM model. While general mixtures of Dirichlet-type distributions are not identifiable on unrestricted parameter spaces, the specific structure of the CDM model (sharing a common mean $\pi$ across components) ensures identifiability for $D \geq 3$ (where $D$ is the number of categories). The constraint $\eta > 1$ further prevents label switching by enforcing an intrinsic ordering where the contaminant component is strictly more dispersed than the reference.

2.3. Estimation and Inference

• Maximum Likelihood Estimation (MLE): Parameters are estimated by numerically maximizing the log-likelihood function. The authors recommend a two-step initialization: first fitting a standard DM model to obtain starting values for $\pi$ and $\sigma$, then using these to initialize the CDM estimation with $\delta \to 0^+$ and $\eta \to 1^+$.
• Model Selection: The Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) are used to compare the CDM model against the standard DM and multinomial models.
• Anomaly Detection: A posteriori probabilities are calculated to classify observations. An observation $y$ is classified as anomalous if the probability of it arising from the contaminant component exceeds 0.5 (though this threshold can be tuned).

3. Key Contributions

1. Novel Distribution: Introduction of the CDM distribution, which explicitly models mild anomalies as a high-dispersion component sharing the same mean as the reference data.
2. Robustness: Demonstration that the CDM model downweights the influence of anomalous observations on parameter estimates, preventing the inflation of the dispersion parameter that plagues standard DM models.
3. Interpretability: The model provides direct interpretations for contamination: $\delta$ quantifies the proportion of anomalies, and $\eta$ quantifies their severity (variability inflation).
4. Automatic Detection: The framework yields a natural rule for outlier detection via posterior probabilities without requiring pre-specification of suspect observations.

4. Results

4.1. Simulation Studies

Two sensitivity analyses were conducted to evaluate the CDM model:

• Single-Point Anomalies: In scenarios where a single anomalous point was introduced at varying distances from the main cluster, the CDM model maintained stable parameter estimates. In contrast, the standard DM model's estimation error increased significantly as the anomaly moved further from the bulk. The CDM model achieved high True Positive Rates (TPR) for distinct anomalies while maintaining low False Positive Rates (FPR).
• Background Noise: In scenarios involving 10% background noise (uniformly distributed and rescaled), the CDM model effectively distinguished between typical and noisy observations. While the TPR did not reach 1.0 (due to some noise resembling typical data), the FPR remained consistently low (< 0.01). The standard DM model showed significant bias in estimating mean proportions, particularly when the noise shifted the data away from the domain center.

4.2. Real-World Application

The model was applied to 16S rRNA sequencing data from Nakatsu et al. (2015), focusing on six bacterial phyla in two subsets: Healthy and Carcinoma (tissue adjacent to tumors).

• Model Fit: In both subsets, the CDM model outperformed the standard DM and multinomial models across all information criteria (AIC and BIC).
• Parameter Estimates:
  • Healthy Subset: The CDM model identified approximately 21.3% of observations as anomalous ($\hat{\delta} = 0.213$) with a high inflation factor ($\hat{\eta} \approx 4.0$). Crucially, the estimated dispersion for the reference component ($\hat{\sigma} = 0.091$) was substantially lower than that of the standard DM model ($\hat{\sigma} = 0.174$), indicating that the standard model was overestimating dispersion due to contamination.
  • Carcinoma Subset: Similar results were observed, with 17.8% of observations classified as anomalous. The reference dispersion was reduced from 0.125 (DM) to 0.068 (CDM).
• Anomaly Identification: The model identified 11 anomalous observations in the healthy subset and 7 in the carcinoma subset. Visual inspection of pairwise scatter plots confirmed that these identified anomalies lay at the periphery of the data cloud, suggesting they represent atypical biological compositions rather than random noise.

5. Significance and Claims

The paper claims that the CDM distribution offers a robust and interpretable tool for analyzing overdispersed multinomial count data, specifically in microbiome research. Its primary significance lies in its ability to:

1. Decouple genuine biological overdispersion from the artificial inflation caused by anomalous observations.
2. Prevent misleading biological inferences that arise when standard DM models attribute contamination to the underlying biological process.
3. Provide a principled framework for anomaly detection that does not require data removal, thereby preserving sample size and statistical power.

The authors conclude that while the model is motivated by microbiome data, its applicability extends to any setting involving multinomial count data where contamination and excess variability are common features. The work does not claim to solve the problem of "gross anomalies" (which require removal) but specifically targets "mild anomalies" through a flexible mixture modeling approach.