[R] se.fit in predict.glm

[(Ted Harding)]

On 27-Apr-04 Peter Dalgaard wrote:
> (Ted Harding) <Ted.Harding at nessie.mcc.ac.uk> writes:
>> The documentation does not say definitely what p$se.fit is,
>> only calling it "Estimated standard errors". I *believe*
>> this means, at each value of X, the SE in the estimation
>> of P[y=1] taking account of the joint uncertainty in the
>> estimation of 'a' and 'b' in the relation
>> 
>>   probit(P) = a + b*X
>> 
>> Can someone confirm that this really is so?
> 
> Pretty accurate, I'd say. 
> 
> Basically, the fitted value is a function of the estimated parameters.
> Asymptotically, the latter are approximately normally distributed with
> a small dispersion so that the function is effectively linear and you
> can approximate the distribution of the fitted value with a normal
> distribution.

Thanks, Peter, that will do nicely! (And spot-on for the
particular application I have in hand).

> Just be aware that the fitted values can be on different scales
> (P vs. logit(P)) and that the se.fit similarly.

I take it your comment refers to the difference, in predict.glm,
between type = "link" (default) and type = "response"?

Thanks, and best wishes,
Ted.


--------------------------------------------------------------------
E-Mail: (Ted Harding) <Ted.Harding at nessie.mcc.ac.uk>
Fax-to-email: +44 (0)870 167 1972
Date: 27-Apr-04                                       Time: 22:50:30
------------------------------ XFMail ------------------------------