Generalised correlations and Simpson's paradox

Pamela J. Davy, J. C. Rayner, Eric J. Beh, V. Pemajayantha, Robert Mellor, M. Shelton Peiris, Jay R. Rajasekera

    Research output: Chapter in Book / Conference PaperConference Paper

    Abstract

    ![CDATA[Simpson’s paradox arises when aggregated data suggest a certain relationship whereas the corresponding non-aggregated data suggest a contrary relationship. Here we relate this well-known paradox to generalised correlations. For a discrete bivariate model with at least one non-binary variable, multiple parameters must be zero in order for independence to hold. These parameters may be expressed in terms of bivariate moments, and interpreted as generalised correlations. Generalised correlations include ordinary Pearson correlation as a special case, but also enable assessment of non-linear relationships. For more than two dimensions, generalised correlations are no longer constrained to sharp bounds of 1 and -1 when linearity holds, so some of the questions accessible in the bivariate case are not accessible in higher dimensions. However generalised correlations provide useful insight into complexities of multivariate structure such as Simpson’s paradox.]]
    Original languageEnglish
    Title of host publicationCurrent Research in Modelling, Data Mining & Quantitative Techniques
    PublisherUniversity of Western Sydney
    Number of pages11
    ISBN (Electronic)0975159909
    ISBN (Print)9780975159903
    Publication statusPublished - 2003
    EventWorkshop on Advanced Research Methods -
    Duration: 1 Jan 2003 → …

    Conference

    ConferenceWorkshop on Advanced Research Methods
    Period1/01/03 → …

    Keywords

    • correlation (statistics)
    • Paradox (computer file)
    • independence
    • multivariate analysis
    • mathematical statistics
    • contingency tables

    Fingerprint

    Dive into the research topics of 'Generalised correlations and Simpson's paradox'. Together they form a unique fingerprint.

    Cite this