Seminario Data: 18/07/07 as 16h Abstract: PERFORMANCE OF OPTIMAL BAYESIAN RANKING METHODS Thomas A. Louis, PhD Department of Biostatistics Johns Hopkins Bloomberg School of Public Health Health services evaluations, environmental assessments, school e©ectiveness studies and identiîcation of active genes depend on the relative position (ranks) of unit-speciîc values. Therefore, ranking to identify high or low rank units (producing league tables) is an im- portant goal. Invalid ranks or inappropriate interpretation can have serious science, policy and înancial consequences. When uncertainties vary over units to be ranked, guidance is needed. For example, basing ranks on hypothesis tests to identify relatively poor perfor- mance unfairly penalizes units with relatively low variance because the tests have higher power; ranking the MLEs unfairly penalizes units with relatively high variance because they tend to be at the extremes. Valid ranking depends on properly melding the order produced by point estimates (MLEs) and the uncertainty of these estimates. Bayesian modeling coupled with loss functions provides the necessary structure. We compare MLE-based ranks, those based on the posterior mean of target parameters and those based on several ranking-relevant loss functions. We evaluate performance, showing that in most realistic situations even optimal methods have limited e©ectiveness and present an application to ranking dialysis providers based on standardized mortality ratios. Clarice --------------------------------------------------------------------- Clarice Garcia Borges Demetrio Departamento de Ciencias Exatas | Phone: +55 019 34294144 R:216 ESALQ/USP | Fax: +55 019 34294346 Caixa Postal 9 | http://ce.esalq.usp.br/ 13418-900 PIRACICABA, SP | BRASIL | clarice@carpa.ciagri.usp.br -------------------------------------------------------------------- 52¦ ReuniÆo Anual da RegiÆo Brasileira da Sociedade Internacional de Biometria e 12§ Simposio de Estatistica Aplicada a Experimentacao Agronomica (SEAGRO) 23 a 27/07/2007, Santa Maria, RS --------------------------------------------------------------------- > > > > Professor Thomas A. Louis, PhD > > > > President of the International Biometric Society > > > > Department of Biostatistics > > > > Johns Hopkins Bloomberg School of Public Health > > > > 615 North Wolfe Street, E3545; Baltimore, MD 21205-2179 USA > > > > 410-614-7838(office); 410-955-0958(fax); 202-494-9331(mobile) > > > > tlouis@jhsph.edu <mailto:tlouis@jhsph.edu> ; http://www.biostat.jhsph.edu/~tlouis/ > > > > > > > > "Biometry, the active pursuit of biological knowledge by quantitative methods" > > > > R. A. Fisher, 1948
Attachment:
ranking-abs.pdf
Description: ranking-abs.pdf