Re: st: twoway graph for imututed values against observed

[weddings@stata.com (Wesley D. Eddings, StataCorp)]

Andy Stoddart (andrew.stoddart@ed.ac.uk) asked about checking the imputations
created by -mi ice-:

> Does anyone know of a way to graph the imputed values over the observed values
> on the same graph?

> For example I am at present using

> mi set wide
> mi ice ? , add(5) [options]

> twoway (hist x1, [options]) (hist _1_x1 if missing(x1), [options]), ///
> legend(label(1 "observed") label(2 "imputed")) 

> I then view each of the 5 imputations by swapping  _1_x1   for
> _2_x1, _3_x1, _4_x1 and _5_x1 in turn to quickly check the imputed
> values appear reasonable

> Obviously it would be preferable to be able to plot all imputed values
> against all observed directly rather than checking individual
> imputations one at a time.

It is possible to use -mi xeq-, rather than repeatedly changing the name of the
variable to be graphed.  Here are the commands to make Andy's plots for all five
imputations:


. generate byte miss_x1 = missing(x1)
. qui mi xeq 1/5: twoway (hist x1 if miss_x1==0) ||   ///
   (hist x1 if miss_x1==1),     ///
   legend(label(1 "observed") label(2 "imputed")); ///
   set more on; more; set more off

It's safe to -generate- the "miss_x1" variable because Andy is using the -wide-
-mi- style.  I took the commands from the note on multiple-imputation
diagnostics, which includes other plots Andy might find useful.  The note is
available at

http://www.stata.com/users/ymarchenko/midiagnote.pdf

It is usual to check each imputed dataset separately.  I do not know of any
research on comparing the aggregate of the observed data to the aggregate of the
imputed values.
   

--Wes
weddings@stata.com
*
*   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/