De-confounding microbiome association studies

Using the largest publically-available 16S dataset of the gut microbiota (from the American Gut Project (AGP)), Ivan Vujkovic-Cvijin and colleagues apply a machine-learning approach to establish the confounding variables of largest effect size in studies that seek to establish associations between the gut microbiota and human disease. They begin by establishing which potential confounders have a strong association with the composition of the gut microbiota, which revealed body-mass index (BMI), sex, age, geographical location, frequency of alcohol consumption frequency, bowel movement quality and dietary intake frequency of various food types. They report significant differences in the distributions of these microbiota-associated variables between cases and controls for most diseases, which suggests that a typical cross-sectional survey would not only identify differences linked to the disease, but also those driven by the microbiota-associated confounding variables unless cases and controls are matched appropriately. Indeed, they show that for many frequent human diseases, matching cases and controls for confounding variables reduces previously observed microbiota differences, and suggest a list of recommended host variables for the purpose of matching comparison groups to reduce the incidence of spurious associations.