Teacher2Teacher Q&A #19164

Outlier detection

T2T || FAQ || Teachers' Lounge || Browse || Search || T2T Associates || About T2T

View entire discussion
[<<prev]

From: Pat Ballew (for Teacher2Teacher Service)
Date: Dec 18, 2007 at 13:19:03
Subject: Re: Outlier detection

Brian, I think you have hit upon essentially the method that is used most often by professional statistical software, except that they use the fact that if the assumptions about the regression data are true, the residuals will be normally distributed and independent of each other. I think Mini-tab reports any residuals that are more than 1.5 or 2 standard deviations from the mean of the residuals (zero should be the mean residual) Using normal deviation, if we assume normal distribution of the residuals, then the IQR is about 1.35 Sigma, and so if you use the Tukey guide (quartile +/- 1.5 IQR) you would have to be somewhere around 2.67 sigma to be an outlier. Keep in mind this is a vertical measurement, and not measuring away from the line horizontally (a slightly smaller distance, I would think). I think you would not be far wrong telling your students to seriously question data with residuals outside 2.5 sigma. HOWEVER>>> I think that the far more important question is to look at the shape of the residuals and the value of r^2 to see if we have the right model. Hope some of this is of help to you. Good luck. -Pat Ballew, for the T2T service

Teacher2Teacher - T2T ®