Seminar by Luke Hagar

Targeted Hypercube Sampling for Scalable Experimental Design

To design non-simplistic studies, we use simulation to explore the suitability of various sample sizes. While these simulations are repetitive, simulation-based design methods can be slow and cumbersome to implement. We propose targeted sampling approaches to drastically reduce the required number of simulation repetitions while maintaining unbiased sample size recommendations. These broadly appliable strategies to improve simulation-based design are illustrated using a method to quickly approximate the power curve for two-group equivalence tests with unequal variances. This application is important because popular free software for equivalence study design does not accommodate the comparison of two groups with unequal variances, and certain paid software solutions that make this accommodation produce unstable results. Our method can be implemented using the dent package in R.