On 3 Jul 2003 05:11:00 -0700, [EMAIL PROTECTED] (Shannon) wrote: > We are trying to pick some predictors for logistic regression. Is > there a rule of thumb on how much variance is needed in order to make > it a possible predictor? Are there other things (besides a priori > assumptions, that is other descriptives) we should look at to > determine what might be a good predictor?
Here is a rule-of-thumb about a predictor that is a dichotomy: if there isn't enough variance so that you can write a 2x2 table that is 'significant', then you won't find a 'significant' contribution. That should hold up pretty well, despite your definition of 'significant'. For continuous variables, the variance represents the range of the sample in hand. You can't expect to do decent prediction for ages 20 to 60 if your data is limited to (20,40). How well can you extrapolate beyond the observed? - that always depend on the R-squared, and it depends in a more subtle way on how much you trust the theory you are modeling. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html "Taxes are the price we pay for civilization." Justice Holmes. . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
