- This event has passed.
Compositional Data Analysis Approach to Body Composition Data
June 2 @ 12:00 pm - 1:00 pm
The AI PHI Speaker Series presents a talk by Curtis Tilves, a Postdoc at the University of Pittsburgh.
Compositional data, or data which are constrained to sum to a whole, lie in a special geometric structure called a simplex. The simplex sample space differs from the real Euclidean space due to this sum constraint; thus, applications of traditional methods (such as correlations or ordinary least squares regression) to simplexed data can result in spurious associations, biased estimates, and inflated type 1 error. However, compositional data analytic methods such as the additive logratio (ALR) transformation can transform data from the simplex space to a real Euclidean space, thus allowing for the application of traditional statistical methods on transformed data.
Body composition should be viewed through a compositional data lens. Body imaging analyses generate images with wholes (total scanned areas) and components (subdivided tissue types). However, these tissue data are often entered into regression models using the untransformed observed values (ex. visceral fat volume) or component ratios (% fat in the trunk), and often to the exclusion of relevant tissues and body size measures. In this talk, I will introduce the ALR method and its importance in models of body composition. While these methods can be applied to any types of imaging data and for any disease, I will demonstrate its application in modeling associations of abdominal and thigh computed tomography scans with diabetes.