Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: estimating standard errors from multiple imputed dataset for single variables - stata10

From   Maarten buis <>
Subject   Re: st: estimating standard errors from multiple imputed dataset for single variables - stata10
Date   Tue, 30 Nov 2010 08:47:54 +0000 (GMT)

--- On Mon, 29/11/10, Deppen, Steve wrote:
> When I use:
> mim, cat(combine)  est(_b[x]) se(_se[x]) : mean x 
> my standard errors seem very small.  I want my SE's to
> reflect the original dataset size and not the dataset X 10.

-mim- is the appropriate command for multiple imputed data. 
However, I would change that command to -mim: mean x-.
It will often(*) result in a reduction of the standard error
compared to techniques that ignore all observations with 
missing values. This is however not due to some artificial 
increase in sample size, actually those extra observations
are used to increase the standard error. 

The idea is that we know something about observations with 
some missing values: the values for those variables where 
that respondent did answer. Multiple imputation is often 
thought of as "creating new values", which is not surprising
given the name. However, it is more useful to thing of 
Multiple Imputation as using the information present in
the observed variables for observations that contains some
missing values. 

The increase in the size of your dataset that you see when
you do multiple imputation is used to take into account that
the there is uncertainty in our imputed values. You obviously
can't use regular statistics on this dataset, as they will 
assume that it is just one large dataset. However, -mim- 
knows that it is a multiple imputed dataset and takes this
into account.

Hope this helps,

(*) Sometimes when the orginal dataset is already large the
gain from using this extra information can actually be smaller
than the uncertainty introduced by using an imputation model,
in which case the standard errors will actually increase.

Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen


*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index