[Paper Review] MedZIM: Mediation analysis for Zero-Inflated Mediators with applications to microbiome data

Zhigang Li, Janaka S. S. Liyanage+10 more|arXiv (Cornell University)|Jun 21, 2019

Gut microbiota and health40 references4 citations

TL;DR

This paper proposes MedZIM, a novel mediation analysis method for zero-inflated mediators in microbiome data, using a two-part potential-outcomes framework to separately model structural zeros and false zeros. It decomposes mediation effects into components driven by zero-inflation and non-zero abundance, improving accuracy over existing methods in simulations and real microbiome studies.

ABSTRACT

The human microbiome can contribute to pathogeneses of many complex diseases by mediating disease-leading causal pathways. However, standard mediation analysis methods are not adequate to analyze the microbiome as a mediator due to the excessive number of zero-valued sequencing reads in the data. The two main challenges raised by the zero-inflated data structure are: (a) disentangling the mediation effect induced by the point mass at zero; and (b) identifying the observed zero-valued data points that are actually not zero (i.e., false zeros). We develop a novel mediation analysis method under the potential-outcomes framework to fill this gap. We show that the mediation effect of the microbiome can be decomposed into two components that are inherent to the two-part nature of zero-inflated distributions. With probabilistic models to account for observing zeros, we also address the challenge with false zeros. A comprehensive simulation study and the applications in two real microbiome studies demonstrate that our approach outperforms existing mediation analysis approaches.

Motivation & Objective

• To address the limitations of standard mediation analysis in handling microbiome data with excessive zeros.
• To disentangle the mediation effect arising from the point mass at zero in zero-inflated distributions.
• To identify and correct for false zeros—observations that are zero due to measurement or technical issues rather than true biological absence.
• To develop a method that integrates probabilistic modeling of zero-inflated data within the potential-outcomes framework for causal mediation analysis.
• To improve the accuracy and robustness of mediation effect estimation in microbiome research

Proposed method

• Proposes a two-part potential-outcomes framework to model the mediation effect in zero-inflated mediators.
• Decomposes the total mediation effect into two components: one driven by the probability of zero (structural zeros), and one by the conditional mean of non-zero values.
• Uses a mixture model to distinguish between structural zeros (true absence) and false zeros (technical or sampling artifacts).
• Incorporates probabilistic models to estimate the likelihood of observing zeros, improving estimation of the non-zero component.
• Applies counterfactual estimands under the potential outcomes framework to identify causal mediation effects.
• Employs likelihood-based inference to estimate parameters and test mediation effects while accounting for the zero-inflated structure

Experimental results

Research questions

• RQ1 How can mediation effects be meaningfully decomposed when the mediator is zero-inflated, especially when the zero component is driven by biological and technical factors?
• RQ2 To what extent can false zeros in microbiome sequencing data bias standard mediation analysis methods?
• RQ3 Can a two-part modeling approach improve the detection of true biological mediation effects in microbiome studies?
• RQ4 How does MedZIM compare to existing mediation methods in terms of Type I error control and statistical power under zero-inflated data?
• RQ5 What is the contribution of zero-inflation versus non-zero abundance to the overall mediation effect in real microbiome datasets?

Key findings

• MedZIM successfully decomposes the mediation effect into components attributable to the zero-inflation process and the non-zero abundance distribution.
• The method demonstrates improved Type I error control and higher statistical power compared to standard mediation methods in simulation studies.
• MedZIM effectively identifies and corrects for false zeros, reducing bias in mediation effect estimation.
• In two real microbiome studies, MedZIM detected significant mediation effects that were missed or underestimated by conventional approaches.
• The decomposition of mediation effects reveals that both the probability of zero and the magnitude of non-zero values contribute meaningfully to disease-related pathways.
• The method maintains robust performance across diverse zero-inflation levels and sample sizes in simulation settings