|Friday, February 22|
|CS06 Theme 3: Prediction and Analytics #2||
Fri, Feb 22, 10:45 AM - 12:15 PM
Average Versus Local EffectsView Presentation Robert Obenchain, Risk Benefit Statistics
*S. Stanley Young, NISS
Keywords: Main Effects, Interactions, Sub-groups, Local Control
Very large medical and environmental data bases are accruing and are being subjected to statistical analysis asking the open-ended question, is some effect present? Now, of course, as n gets large the standard error of the mean gets small so statistical significance is a forgone conclusion. People respond differently to medical interventions and to environmental exposures, so average effects can be meaningless and very small, statistically significant average effects most likely are meaningless. One size does not fit all. What to do? The idea is to use statistical methods that are designed to deal with heterogeneous human response. Recursive partitioning can be used to find subsets of people that respond in their own way. Clustering of a large data set and looking for unique responses within clusters, local control, is another way to examine individualized responses. Recursive partitioning and local control analysis on both simulated and real data sets will be presented. The benefit of these analysis methods is that we can move away from claims based on potentially meaningless average effects.