Scribd Logo
Upload

Welcome to Scribd!

  • 17 views

    Regression Outliers

    Uploaded by

    Ross Zhou
    statistics

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Regression Outliers

    Uploaded by

    Ross Zhou
    17 views2 pages
    statistics

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    statistics

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    17 views2 pages

    Regression Outliers

    Uploaded by

    Ross Zhou
    statistics

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 2

    REGRESSION OUTLIERS

    1. Identification of Outliers
    An outlier is an extreme observation. Typically points further than, say, three or four
    standard deviations from the mean are considered as outliers. In regression however,
    the situation is somewhat more complex in the sense that some outlying points will have
    more influence on the regression than others. In JMPIN there is one diagnostic that can
    be used to identify possibly influential outliers, known as Cooks Distance, or simply
    Cooks D. Given a regression of Y on ( x1 ,.., xk ) using data set ( y j , x1 j ,.., xkj ), j = 1,.., n ,

    if
    s =

    estimated root mean square error,

    y j =

    regression estimate of the conditional mean E (Y j | x1 j ,.., xkj ) ,

    y j (i ) = regression estimate of the conditional mean E (Y j | x1 j ,.., xkj ) with the


    i th data point ( yi , x1i ,.., xki ) removed,

    then Cooks Distance for point i is given by

    ( y
    D =
    n

    j =1

    y j (i ) )

    (k + 1) s 2

    , i = 1,.., n

    Intuitively, Di is a normalized measure of the influence of point i on all predicted mean


    values, y j , j = 1,.., n . Cooks D can be obtained using Fit Model in JMPIN as follows:
    (i) Right click on the heading of the Parameter Estimates table,
    (ii) Select the Save Columns options, and click on Cooks D Influence.
    (iii) A new data column will appear, contain the Cooks D Influence values.
    To identify potential outliers, one Rule of Thumb is to treat point i as an outlier when:

    Di

    4
    n (k + 1)

    As with all Rules of Thumb, this provides only a rough guideline (and often tends to
    identify too many points as potential outliers). The best strategy is to look at the
    distribution of Cooks D values and see whether there are any conspicuously large values
    relative to the others. If these values are roughly of the magnitude 4 /(n k 1) or larger,
    then they are worth investigating further.
    2. Treatment of Outliers

    The key point to stress here is that the above procedure can only serve to identify points
    that are suspicious from a statistical perspective. It does not mean that these points should
    automatically be eliminated! The removal of data points can be dangerous. While this
    will always improves the fit of your regression, it may end up destroying some of the
    most important information in your data.
    Hence the first question that should be asked is whether there exists some substantive
    information about these points that suggests that they should be removed. Do they
    involve special properties or circumstances not relevant for the situation under
    investigation? Do they involve possible measurement errors? If no such distinguishing
    features can be found, then there are no clear grounds for eliminating outliers.
    An alternative approach is to perform the regression both with and without these outliers,
    and examine their specific influence on the results. If this influence is minor, then it may
    not matter whether or not they are omitted. On the other hand, if their influence is
    substantial, then it is probably best to present the results of both analyses, and simply
    alert the reader to the fact that these points may be questionable.

    You might also like