Mediation analysis for zero-inflated mediators with applications to microbiome data
Zero-inflated data is commonly seen in biomedical research such as microbiome studies and single-cell sequencing studies where zero-valued sequencing reads arise due to technical and/or biological reasons. Mediation analysis approaches for analyzing zero-inflated mediators are still lacking largely because of challenges raised by the zero-inflated data structure: (a) disentangling the mediation effect induced by the point mass at zero is not straightforward; and (b) identifying the observed zero-valued data points that are actually not zero (i.e., false zeros) is difficult. Existing approaches for analyzing microbiome as a high-dimensional mediator can not handle the zero-inflated data structure. In this paper, we developed a novel mediation analysis method under the potential-outcomes framework to fill the research gap. We show that the mediation effect of a zero-inflated mediator can be decomposed into two components that are inherent to the two-part nature of zero-inflated distributions: the first component corresponds to the mediation effect attributable to a unit-change over its positive domain (often counts or continuous values) and the second component corresponds to the mediation effect attributable to discrete jump of the mediator from zero to a non-zero state. With probabilistic models to account for observing zeros, we can also address the challenge with false zeros. A comprehensive simulation study and the applications in two real microbiome studies demonstrate that our approach outperforms current standard mediation analysis approaches.
READ FULL TEXT