On Wed, 27 Jan 2010 09:30:45 -0800, Jeremy Miles <[log in to unmask]> wrote:
>You can standardize (or divide by 2 SDs - Gelman recommends that in a
>paper somewhere, and in his regression book), or somehow equate the
>scales. Or you can try to make a sensible comparison - e.g. in one
>analysis I did, I said that smoking is equivalent to being 8 years
>older, in terms of the odds of having some condition (I think it was
>Dupuytrens).
>
>However, if you standardize, you throw away the original information
>about the scales, which can be a problem, because your comparison is
>sensitive to the SD of the data in your sample.
I'm not a fan of standardization for these reasons, but Gelman's position seems to be to
semi-standardize by just rescaling the predictors as a default in this way. This is less
problematic than also standardizing the outcome variable. An equivalent scaling is
obtained by using effect coding for categorical predictors and dividing by 1 SD.
Another scaling is to use percentage or proportion of maximum performance (POMP). This
doesn't get distorted by the bias in estimating the SD or by reliability etc. However, it
requires that you know the min and max possible.
However, this doesn't really get around the problem of comparability - which is more of a
philosophical problem. I'd consider reversing it so you can see how much change in a
continuous predictor is required to produce the same OR change as your dichotomous
variables. This will get clumsy of you have lots of dichotomoous predictors.
Best of all you can try and graph the effects - as a good plot can often reveal the
complexity in the data.
Thom
|