A Rigorous Framework for Automated Design Assessment and Type I Error Control: Methods and Demonstrative Examples (Working Paper)
share
We present a proof-by-simulation framework for rigorously controlling a trial design's operating characteristics over continuous regions of parameter space. We show that proof-by-simulation can achieve two major goals: (1) Calibrate a design for provable Type I Error control at a fixed level alpha. (2) Given a fixed design, bound (with high probability) its operating characteristics, such as the Type I Error, FDR, or bias of bounded estimators. This framework can handle adaptive sampling, nuisance parameters, administrative censoring, multiple arms, and multiple testing. These techniques, which we call Continuous Simu lation Extension (CSE), were first developed in Sklar (2021) to control Type I Error and FWER for designs where unknown parameters have exponential family likelihood. Our appendix improves those results with more efficient bounding and calibration methods, extends them to general operating characteristics including FDR and bias, and extends applicability to include canonical GLMs and some non-parametric problems. In the main paper we demonstrate our CSE approach and software on 3 examples (1) a gentle introduction, analyzing the z-test (2) a hierarchical Bayesian analysis of 4 treatments, with sample sizes fixed (3) an adaptive Bayesian Phase II-III selection design with 4 arms, where interim dropping and go/no-go decisions are based on a hierarchical model. Trillions of simulations were performed for the latter two examples, enabled by specialized INLA software. Open-source software is maintained at (1)
READ FULL TEXT