Statalist The Stata Listserver

[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: RE: Test for missingness type

From   "Maarten Buis" <[email protected]>
To   <[email protected]>
Subject   st: RE: Test for missingness type
Date   Wed, 27 Sep 2006 11:51:45 +0200

--- m�hlmann wrote:
> I have a dataset with one variable (x) partially missing.  I want to
> test if the missingness mechanisms is nonignorable (data is MNAR), ie if
> P(r=1|x,z)  does depend on x, where r is the binary missingness
> indicator and z are completely observed covariates.

Short answer is: "you can't test the MAR/NMAR assumption".

Slightly longer answer is: 
You want to test whether or not x is missing depends on x itself. 
In order to test that you would need to compare the values of x 
when x is observed and the values of x when x is missing. So you
would need to know the values of x when x is missing. If you would 
know that, x wouldn't be missing... So MAR is an untestable 

A common (and in my humble opinion reasonable) way around this
is to assume that missingness on x may depend on lots of other
(fully or partially observed) variables, and assume that after controlling
for these other variables there is nothing left for x itself, i.e. the MAR
becomes more reasonable when you control for more covariates. 


Maarten L. Buis
Department of Social Research Methodology 
Vrije Universiteit Amsterdam 
Boelelaan 1081 
1081 HV Amsterdam 
The Netherlands

visiting adress:
Buitenveldertselaan 3 (Metropolitan), room Z434 

+31 20 5986715

*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index