Abstract:
|
Sparse data often result from epidemiologic research due to a combination of multiple factors such as low event rate or rare exposure. Regression estimates of such data are known to be biased and/or imprecise. Application of Bayesian regression strengthens sparse data and are said to produce more reliable estimates. This research was motivated from an analysis exploring the effect of lipids on myocardial infarction (MI) in Japanese females, among whom MI occurrence and prevalence of smoking are rare, resulting in sparse data. Using logistic regression, we will compare different Bayesian regression-fitting methods (Markov chain Monte Carlo (MCMC) or approximation using maximum likelihood method or data augmentation priors) as well as the effect of different priors (Jeffrey's invariant prior, weakly informative priors, and highly informative priors) on sparse data analysis through simulation studies and apply them to motivational data analysis. We will show the weakly informative priors with MCMC-fitting can effectively shrink estimates and estimated interval length obtained from sparse data towards background medical knowledge, while it is not influential to non-sparse variables.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.