419 – Contributed Oral Poster Presentations: Section on Physical and Engineering Sciences
Evaluation of a New Screening Procedure
Steve Webb
The performance of a sequential screening procedure for finding a few active factors from a larger number is investigated. Called the Winnow, it was previously introduced and studied using dummy simulation models. Here we apply it to a legacy simulation using eight different baselines with from 25 to 78 inputs varied. Large fractional factorials are used to find "true" values of the effects. The largest were of the order of one standard deviation for these baselines. The sequential procedure is evaluated by finding how many real effects are identified and the accuracy of the resulting estimates. It is shown that an effect as large as 1 sigma will very likely be found and one as large as 0.5 sigma will usually be found. The winnow is found to be generally better than the alternatives of fixed supersaturated designs or a version of group screening.