On 16 Mar 2004 20:13:31 -0800
[EMAIL PROTECTED] wrote:
> On 16 Mar 2004 at 12:26, Phillip Good wrote:
>
> > I was unaware that maximum likelihood had any desirable properties
> > except in the case of normally-distributed random variables where the
> > max likelihood approach leads to estimators that are desirable for
> > entirely different reasons.
> >
>
> Could you please explain what in your opinion is wrong with likelihood
> methods, which in effect makes up the workhorse of todays applied
> statistics, not only for normal models, but for instance in generalized
> linear models and a lot of others?
To me, the point is that the maximum likelihood methods that are used in
missing data situations make too many assumptions. One of the
assumptions that are made too quickly with such models is that of
linearity, not to mention conditional normality.
When I do multiple imputation I frequently need to use nonlinear
imputation models.
---
Frank E Harrell Jr Professor and Chair School of Medicine
Department of Biostatistics Vanderbilt University
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
. http://jse.stat.ncsu.edu/ .
=================================================================
