iSixSigma

Outliers – Binary Logistic Regression Event Probability

Six Sigma – iSixSigma Forums General Forums Methodology Outliers – Binary Logistic Regression Event Probability

Viewing 4 posts - 1 through 4 (of 4 total)
  • Author
    Posts
  • #53506

    Doa
    Participant

    Hi to All,

    Does the term outlier have use in the following instance?

    As a result of running Step wise and Binary Logistic Regression I obtained one significant predictor or “X” of my response “Y”. I have plotted the event probability of “Y” versus “X” all in Mintab, all apears as expected except for one or two datapoints out of 30 total have event probabilites double that of the surronding datapoints.

    Is calling these datapoint Outliers accurate? If so, are they handled the same way?

    I hope this is clear. I’ve attached a copy of this in Excel.

    Thank you in advance….

    Regards,
    Marty

    0
    #190409

    Doa
    Participant

    Here is the file uploaded as a word file…….. [file name=EPRO_Ex.docx size=28697]https://www.isixsigma.com/images/fbfiles/files/EPRO_Ex.docx[/file]

    0
    #190410

    Robert Butler
    Participant

    In binary regression this issue is that of influential observations. I suspect somewhere in the Minitab program you have the option of computing the influence statistics ( usually called DFBETAs) for the model and you probably have the option of plotting these against the estimated outcome probabilities (it may be that you have done this and this is what you mean when you say you plotted event probabilities of Y vs. X).

    The usual procedure is to rerun the model with the high DFBETA points eliminated and then check to see if their removal makes a noticeable change in the model coefficients. If there isn’t a noticeable changed then there is no reason to suspect the data points in question and you can use the model built using all of the data points.

    0
    #190414

    Doa
    Participant

    Hi Robert,

    Thanks for the response….

    Regards,

    Marty

    0
Viewing 4 posts - 1 through 4 (of 4 total)

You must be logged in to reply to this topic.