Imprimir

Regression methods for multiple outcomes in health research

  • Main research unit: Centro de Investigação em Tecnologias e Sistemas de Informação em Saúde (CINTESIS/FM/UP)
  • Principal Investigator:  Este endereço de email está protegido contra piratas. Necessita ativar o JavaScript para o visualizar.
  • Researchers:
    • Jaroslaw Harezlak
    • Eric Tchetgen Tchetgen
  • Start date: 01.03.2009 | End date: 30.09.2012
  • Financing: € 59,938.00

Project description: In Health research it is typical to formalize the research question in terms of a single, unidimensional outcome. Even when the problem is intrinsically multidimensional, resulting on the collection of multiple outcomes, the most common strategy involves summarizing the outcomes into a single score using some pooling strategy or, alternatively, analyzing each outcome separately using univariate methods [1-4]. These approaches are practical, both from the point of view of comprehension/interpretation of the results and simplification of the analysis.

However, in many situations it is not possible to simplify the research question to a one-dimensional outcome (e.g. [5-8]) and the separate analysis of each outcome does not consider the problem as whole, Even in the context of clinical trials, where the topic of multiple outcomes received some attention due to the increasing number of trials with multiple endpoints, the main discussion has focused on adjustments to the significance level for multiple testing due to the individual analysis of the endpoints  The MRS study described in the "Research Plan section" is a prime example where interest lies on studying the association of several factors with 12 outcomes: 4 metabolite ratios in 3 brain regions. If the multivariate structure of the data is ignored, and each outcome is analyzed separately, we may lose the global understanding of the relationships between the factors and the outcomes.

Despite the existence of multivariate methods that can be used for the analysis of multiple outcomes, the reality is that these methods are hardly found in applied research and do not apply to many settings. For example, the multivariate regression analysis of non-commensurate outcomes (outcomes measured on different scales or mixed type of outcomes such as continuous and binary outcomes) still lacks a complete theoretical framework. The development of appropriate methodology for model selection in the multivariate setting is also seriously lacking.

In this project we propose to develop a general framework for the analysis of Health related data involving multiple outcomes, including the general situation of non-commensurate outcomes and model selection. It would be unrealistic to anticipate that we will solve the entire problem of handling multiple outcomes (for example, we do not consider longitudinal or clustered data) but we believe that we will be able to establish the foundation for other methodological extensions. In any case, we foresee that such methodology could be applied to a variety of examples, in many areas of Health research. In particular, the project is motivated, in part, by a real data example: the MRS study of long term brain damage in HIV patients, carried out by the HIV Neuroimaging Consortium at multiple sites in USA.

The three Biostatisticians involved in this proposal got to know each other at the Harvard School of Public Health during their PhD studies. However, their interest in this topic grew independently, mainly motivated by real data problems in their practice of collaborative research. Their individual methodological research interests and expertise provide an excellent combination for the success of this project. Armando Teixeira Pinto has been working on multivariate methods for non-commensurate outcomes [see references 1 and 2 of the “Past publications"], Eric Tchetgen Tchetgen is interested in robust and efficient estimation [see references 3 and 5 of the “Past publications”] and Jaroslaw Harezlak has been studying model selection methods [see references 4 and 5 of the “Past publications"].

Although this is a fairly technical and quantitative project, its results have an important and direct implication in problems related to Epidemiology, Public Health, Health Technology Assessment and Health Services Research that fully justify the submission of this project

Secção Ensino Secção investigacao Secção consultadoria
   Mestrados | Doutoramentos | Formação Contínua         Projetos | Teses e Dissertações | Publicações        Bioestatística | Informática Médica