Re: logistic model effect range

Richard Ulrich wrote:
> On Fri, 26 Aug 2005 07:32:51 -0600, Trevor Wiens
> <twiens@xxxxxxxxxxxxxxxxxxxxx> wrote:
>
>
>>I have a logistic regression model where one of the variables (which is
>>continuous) has a effect range that crosses zero. The model was created
>>using R, from which I can produce table which provides predictor variable
>>details such as low, high,range, effect, S.E., lower 0.95 and
>>upper 0.95 (for both the variable and its odds ratio).
>>
>>If the range between the lower and upper 0.95 crosses zero, (or for the
>
>
> If the range of the 95% CI crosses zero, that *illustrates*
> the simple fact that the p-level is not less than 0.05; the
> CI is typically obtained by inverting an equation with
> the p-value.
>
>
>>odds ratio crosses 1) this would indicate to me I'm seeing a non-linear
>>effect, such as a preference for moderate values, etc. BTW, I've noticed
>>that in cases where this occurs, the P value for that predictor variable
>>is always very large (not significant).
>
>
> There is not any new information in the Confidence
> Interval, concerning nonlinearity or the conceivable use
> of moderate values.
>
>
>>Can someone confirm that my understanding is correct on this? If this is
>>the case, should that predictor variable be removed from the model? Are
>>there methods to transform the variable so that it can still be used?
>
>
>
> The Mosteller-Tukey book on regression gives a good
> introduction to transforming variables. Experimental
> design is important in the problem of when it is acceptable
> to drop variables. Using the p-level as your only guide is
> generally considered bad -- You can google groups
> < group:sci.stat.* stepwise > or look in my stats-FAQ for
> old comments.
>
> Hope this helps.


I would add to Rich's comment, since you are using R, that you might
want to consider Frank Harrell's excellent book "Regression Modeling
Strategies", which includes examples in SAS and S-PLUS/R.

Frank has kindly contributed his Design and Hmisc packages to R
(available from CRAN) which aid tremendously with this domain.

More information is here:

http://biostat.mc.vanderbilt.edu/twiki/bin/view/Main/RmS

HTH,

Marc Schwartz
.