Re: st: RE: St: Panel data imputation

[David Bai <db555@mail.com>]

Thanks, Nick. This makes good sense. Thanks again to Nick and Maarten 
for the help.


-----Original Message-----
From: Nick Cox <n.j.cox@durham.ac.uk>
To: statalist@hsphsun2.harvard.edu <statalist@hsphsun2.harvard.edu>
Sent: Tue, Sep 21, 2010 12:32 pm
Subject: RE: st: RE: St: Panel data imputation


How we can tell?

Why not do both and see what difference it makes and show your readers?

Nick
n.j.cox@durham.ac.uk

David Bai

Thanks Nick and Maarren.
My impression is that, ignore missing values (default approach in
Stata), which I assume is listwise approach, has been critisized by
many researchers, such as Paul Allison, because the sample without
missing values may end up to be very different from the original
population. So whether the results derived from the listwised cases can
be generalizable to the original population is suspectable. That's why
methods like MI have been developed. Although MI has its limitations, I
assume that it is better than using listwise-deleted sample that no
longer represent the original population?

Maarten buis <maartenbuis@yahoo.co.uk>

--- On Tue, 21/9/10, David Bai wrote:
> I have many more variables that can be possibly related to
> revenue. Given what you and Maaren explained below, I guess
> using ipolate and year info only might not be an accurate
> way to predict revenue. MI might be a better approach.
> Correct me if I am wrong. Thank you.

That really depends on all gritty little details of your
data analysis: what you want to do with your imputed data, why
some variables have missing values, what assumptions you are
willing to make, etc. etc. So, the not very helpful "correct"
answer is that "it depends". In general I would recommend to
just ignore missing values (default approach in Stata). Methods
like MI are great but also very sensitive and hard to diagnose,
so unless you really know what you are doing I would stay away
from those techniques.

Think of it this way: I will generaly be skeptical about results
when MI noticably changes it. This might legitemately happen, but
it more often points into the direction of an error in your
imputation model. Now, why would I go through the effort of learning
about MI if I am only going to believe the results when the MI does
not change them?

There are specific situations where MI make perfect sense, but
MI is not suitable as a default. The problem is, that making such a
decision depends on all the gritty little details of your project,
your theory, your research question, your discipline, and much more.
So, this is really a decision that you will have to make on your own.


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

 

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/