Title: Multiblock analysis of mixed data with optimal scaling: Application in epidemiology
Authors: Martin Paries - Oniris (France) [presenting]
Evelyne Vigneau - National College of veterinary medicine, food science and engineering (France)
Stephanie Bougeard - ANSES (France)
Abstract: A common problem in modern science, especially in biology, is the exploration of the relationships between blocks of variables measured on the same observations. Unsupervised component-based multiblock methods, such as Generalized Canonical Correlation Analysis or Multiblock Principal Component Analysis, are well referenced and allow for exploring these relationships. However, they are designed for numeric data only and real data have actually various formats (i.e nominal, ordinal, numerical = mixed data). Several component-based methods are proposed in the literature to deal with mixed variables, the well-known ones pertaining to the framework of Optimal Scaling. More precisely, Optimal Scaling is based on the two-step ALSOS algorithm where the Optimal Scaling step (i.e quantification of variables) alternates with the least square estimation step of the model parameters. Within this Optimal Scaling context, we propose an exploratory multiblock method called Multiblock Principal Component Analysis with Optimal Scaling (MBPCAOS). The proposed MBPCAOS can mainly be compared to other ones in the same framework, such as MFAmix or Overals. The MBPCAOS method is illustrated in a real case study pertaining to epidemiology.