Email Sanne Willems
Personal academic webpage
New Approaches in Survival Analysis
Optimal scaling with regularization and survival analysis are two research elds in statistics. In the history of the methodology for the social and behavioral sciences, there has been a strong demand for approaches that deal with categorical data. Classical statistical methods had to be adapted to suit particular characteristics of research in, for example, psychology, education, political science, and market research. These adaptations were aimed at the optimal assignment of quantitative values to qualitative scales, and have been actively developed in the area of psychometrics.
In the medical sciences, there is ample opportunity of application of optimal scaling as well. In survival analysis, the predictor is usually a composite variable derived from a large number of categorical variables, for example measuring depression. The basic data are non-numerical, with measurements recorded on scales having an uncertain unit of measurement. Such qualitative or categorical variables describe the objects (patients) in a limited number of categories. Ignoring the particular characteristics of the data, the typical approach in survival analysis it to simply compute a sum score as a composite variable. The incorporation of optimal scaling in survival analysis seems to have been given very little attention. This situation motivates the PhD research described in this proposal. Quite recently optimal scaling has been combined with regularization to improve the prediction accuracy of regression models. Our ultimate aim is to combine the two elds to improve the prediction accuracy of the proportional hazards model.
In the sequel, I will rst brie y describe the basics of classic survival analysis and introduce notation. Then, I will describe the data set that will be used, being an example of a typical data set used in survival analysis. This data example will show the relevance of the envisaged research, since it will become clear that optimal scaling is extremely appropriate for the type of data under consideration. Finally, a time line will be given which gives an impression of my activities during my PhD.
Prof. Dr. J.J. Meulman & Dr. M. Fiocco
1 September 2014 – 1 September 2018