Session IX (May 17, 8:30am-10:00am): Optimal Experimental Design, organized by Abhyuday Mandal
Title: A Supersaturated Screening Design Framework Based on Lasso Support Recovery
Speaker: Kade Young, Eli Lilly
Abstract: Screening experiments utilize n experimental runs to determine which of $p$ factors drive a response. Supersaturated screening designs (SSDs) represent a screening experiment where $n<p+1$. In this case, the full main effects linear model cannot be uniquely estimated with Ordinary Least Squares. Thus, it is common for some type of penalized estimation method, like the lasso, to be used to perform factor screening. This works develops a framework for optimal SSDs based on maximizing the support recovery probability of the lasso and shows that a compound symmetric matrix (a matrix where all off-diagonals are equal) is the ideal structure of lasso information matrices for support recovery. This ideal structure allows for the theoretical justification on why some two-level SSD criteria out-perform others under certain assumptions about the signs of the effects. Additionally, a design construction algorithm is presented that balances two design criteria based on how close a given design's information matrix is to the ideal structure. The performance of these constructed designs are evaluated compared to existing SSD construction methods, and their utility and limitations are discussed.