A1249
Title: Design and analysis of a new class of three-level factorial design for screening experiments
Authors: Cinyu Shih - National Taiwan University (Taiwan) [presenting]
Frederick Kin Hing Phoa - Academia Sinica (Taiwan)
Abstract: Three-level screening designs aim to estimate main effects and second-order terms under limited experimental runs while minimizing aliasing and capturing nonlinear behavior. However, widely used designs such as definitive screening designs (DSDs) and orthogonal-minimal-aliasing response surface (OMARS) designs rely fully or partially on search-based or optimization-driven constructions. This reliance increases computational costs, limits scalability, and restricts flexibility. Moreover, changing factor sets often necessitates a full reconstruction. To address these limitations, a new class of designs is proposed, termed OMAHLD, constructed via a closed-form, search-free procedure. The proposed design accommodates both odd and even numbers of factors, scales naturally to high-dimensional settings, and allows sequential expansion by adding columns with minimal increases in run size. OMAHLD preserves the minimal-aliasing principle, and any residual aliasing can be efficiently resolved using simple matrix operations. To further mitigate main-effect aliasing, a two-stage Dantzig selector tailored to OMAHLD designs is developed. Simulation studies demonstrate that this strategy achieves true-positive and false-discovery rates comparable to those of DSDs, while substantially outperforming conventional one-stage methods. These results establish OMAHLD as a flexible and high-performance alternative for large-scale three-level screening experiments.
