 |
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
st: suest and micombine
Would it be reasonable to use bootstrapping in this case? The algorithm
I have in mind is:
1. bootstrap sample
2. run each regression for each imputation (i.e., regress ... if _mj==1)
3. average the coefficients for each regression over the imputations 4.
calculate the difference between the coefficients of interest 5. test if
the distribution of differences created by bootstrapping contains zero
(to whatever confidence level you like) If the distribution is not
sufficiently normal, you could try bias-corrected and accelerated
confidence intervals.
I'm offering this as a suggestion, but also wondering if others out
there think this is a good approach. I've thought about this because I'm
in need of a similar solution.
Maggie
------------------------------
Date: Thu, 19 Feb 2009 10:38:17 +0000 (GMT)
From: Maarten buis <[email protected]>
Subject: Re: st: suest and micombine
- --- On Thu, 19/2/09, [email protected] wrote:
> I want to test whether b coefficients vary between groups.
> I first store estimates after running "micombine regress" over several
> imputations.
> However, after running -suest- I get message "option res not allowed"
> (although no options are specified).
The thing to remember about multiple imputation is that it is primarily
designed to make inference about coefficients in isolation. Making
inference about combinations of coefficients in a multiple imputation
context is going to be very hard, though some progress has been made,
see for example this talk by Rose Medeiros at the 2008 Fall North
American Stata Users' Group Meeting:
http://ideas.repec.org/p/boc/fsug08/11.html .
Making cross model inference is even harder, and I wouldn't be surprised
if that hasn't even been worked out in theory, let alone implemented in
software.
- -- Maarten
*
* 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/