Análise de sobrevivência com erros de classificação desconhecidos

• Ano de defesa: 2013
• Autor(a) principal: Raquel de Souza Borges Ferreira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Erros de classificação; Análise de sobrevivência; Subconjunto de validação; Análise de sobrevivência (Biometria); Estatistica
• Link de acesso: http://hdl.handle.net/1843/ICED-9H4FS3

When the interest resides in studying the time to an event, and the detection of this event is conditional to the outcome of some test that, in turn, may be subject to misclassification, the quality of the imperfect test is incorporated through its measures of sensitivity and specificity. However, when there is no information on such parameters, the model presents a problem of non-identifiability. This thesis aims at evaluating and solving the problem of non-identifiability using a Bayesian approach, and in addition analyzing the impact of restrictions on prior distributions for the parameters of sensitivity and specificity of the imperfect diagnostic test. Despite the increase of information via the restrictions, for relatively large samples posterior distributions are not sensitive to them. Validation subsets are also incorporated. In a study of a dataset used as example which has gold standard outcomes to all individuals, it was verified that the model with a validation subset can estimate satisfactorily the parameters. Moreover, simulation studies show improvements by incorporating such subsets.