[R] Pointwise Confidence Bounds on Logistic Regression
Prof Brian Ripley
ripley at stats.ox.ac.uk
Wed Jun 18 23:32:48 CEST 2008
On Thu, 19 Jun 2008, Rolf Turner wrote:
>
> On 19/06/2008, at 8:08 AM, Bryan Hanson wrote:
>
>> Hi all. I hope I have my terminology right here...
>>
>> For a simple lm, one can add “pointwise confidence bounds” to a fitted line
>> using something like
>>
>>> predict(results.lm, newdata = something, interval = "confidence")
>>
>> (I'm following DAAG page 154-155 for this)
>>
>> I would like to do the same thing for a glm of the logistic regression
>> type,
>> for instance, the example in MASS pg 190-192 (available in the help page
>> for
>> predict.glm).
>>
>> However, predict.glm does not have the same kind of features as "plain old"
>> predict, i.e. One cannot specify interval = "confidence"
>
> I guess that one reason for that is that prediction intervals
> rarely if ever make sense with generalized linear models. So only
> one kind of interval is in effect possible.
>>
>>> From what I've read, "pointwise confidence bounds" are computed from the
>> SE's for each point. However, I don't see quite where to extract this
>> information with a glm
>>
>> So, is there an existing function that does what I am describing for a glm,
>> or can someone point me in the right direction to start writing my own?
>
> Use predict(<whatever>,type="response",se.fit=TRUE). You get a list with
> three components, the first two of which are the fitted values and their
> standard errors. (The third is the ``scale'' factor, usually/often equal to
> 1.)
I would suggest rather computing confidence intervals on linear predictor
scale and transforming those to response scale. That way you will not get
e.g. negative values for probabilities in a logistic regression.
>
> cheers,
>
> Rolf Turner
--
Brian D. Ripley, ripley at stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272866 (PA)
Oxford OX1 3TG, UK Fax: +44 1865 272595
More information about the R-help
mailing list