Multiple correspondence discriminant analysis: An application to detect stratification in copy number variation

Authors

  • Alejandro Caceres,,

    1. Center for Research in Environmental Epidemiology (CREAL), Parc de Recerca Biomedica de Barcelona, 88 Doctor Aiguader, Barcelona, Spain
    2. Institut Municipal d'Investigació Mèdica (IMIM), Parc de Recerca Biomedica de Barcelona, 88 Doctor Aiguader, Barcelona, Spain
    Search for more papers by this author
  • Xavier Basagaña,

    1. Center for Research in Environmental Epidemiology (CREAL), Parc de Recerca Biomedica de Barcelona, 88 Doctor Aiguader, Barcelona, Spain
    2. Institut Municipal d'Investigació Mèdica (IMIM), Parc de Recerca Biomedica de Barcelona, 88 Doctor Aiguader, Barcelona, Spain
    3. CIBER Epidemiología y Salud Pública (CIBERESP), Parc de Recerca Biomedica de Barcelona, 88 Doctor Aiguader, Barcelona, Spain
    Search for more papers by this author
  • Juan R. Gonzalez

    Corresponding author
    1. Center for Research in Environmental Epidemiology (CREAL), Parc de Recerca Biomedica de Barcelona, 88 Doctor Aiguader, Barcelona, Spain
    2. Institut Municipal d'Investigació Mèdica (IMIM), Parc de Recerca Biomedica de Barcelona, 88 Doctor Aiguader, Barcelona, Spain
    3. CIBER Epidemiología y Salud Pública (CIBERESP), Parc de Recerca Biomedica de Barcelona, 88 Doctor Aiguader, Barcelona, Spain
    • Center for Research in Environmental Epidemiology (CREAL), Parc de Receerca Biomedica de Barcelona, 88 Doctor Aiguader, Barcelona, Spain
    Search for more papers by this author

Abstract

We illustrate the use of multiple correspondence analysis (MCA) to correct for population stratification of copy number alteration data. In addition, we propose the use of multiple correspondence discriminant analysis (MCDA) to identify an optimal set of copy number variants (CNVs) that correctly infers the population stratification of a CNV map. Within MCDA, we highlight the novel use of correlation with class directions for variable ranking. We found a set of 20 CNVs with 98 per cent predictability in a CNV map of the HapMap populations. On this sample, the selection of variables based on centroid ranking outperformed the most common practice of ranking variables with their correlation to the principal axes. Copyright © 2010 John Wiley & Sons, Ltd.

Ancillary