Re: st: Using -estout- with multiply-imputed data: how to report r2 and r2_a

[ymarchenko@stata.com (Yulia Marchenko, StataCorp LP)]

Nic <3817629@hotmail.ca> uses the user-written command -mibeta- with another
user-written command -estout- and is having trouble accessing some of the
estimation results:

> I am trying to create some tables using -estout- which compare several OLS
> models. I'd like to include 'R-Square' and 'Adjusted R-Square' in the
> summary statistics. Acock (A Gentle Introduction to Stata, 2010) says to use
> the -mibeta- command written by Julia Marchenko. This does indeed produce
> the results in Stata's screen output, but I cannot figure out how to get the
> -estout- package to include these summary statistics in its output. Here is
> my .do file (as used on a previously multiply-imputed data set):
>
> -----
> eststo clear
> eststo: quietly mibeta d1 g c* s*, fisherz miopts(post)
> eststo: quietly mibeta d1 g c* s* f*, fisherz miopts(post)
> eststo: quietly mibeta d1 g c* s* f* e*, fisherz miopts(post)
> esttab using table7.csv, cells("b(fmt(3)) se(fmt(2)) p(fmt(3)star)") 
> drop(_cons) noobs stardetach starlevels(* 0.05) nonumbers mlabels("Model 1" 
> "Model 2" "Model 3") stats(r2 r2_a, labels("R-Square" "Adjusted R-Square"))
> ----
> I think I have figured out that -eststo- will not work with -mibeta- unless
> the 'miopts(post)' option is used. Without this option, my tables would not
> include any data at all (i.e., the b, se, and p would all be blank). I can
> now produce a table with b, se, and p for all the models, but the r2 and
> r2_a are still blank. I believe this means that they are not being 'posted'?

-mibeta- is a wrapper for -mi estimate: regress-.  It does not save any
additional estimation results other than those saved by -mi estimate:
regress-; see "Saved results" in -help mibeta-.  The -mi estimate- command
does not store matrices -e(b)- and -e(V)- by default.  Multiple-imputation
(MI) estimates and their VCE are stored in -e(b_mi)- and -e(V_mi)-,
respectively.  You can use the -post- option with -mi estimate- to post
multiple-imputation results to -e(b)- and -e(V)-.  As Nic already found out,
this option can be specified with -mibeta- as -miopts(post)-.  Scalars -e(r2)-
and -e(r2_a)- are not saved by -mi estimate- and thus cannot be accessed after
-mibeta-.

> So my question is: is it possible to use the -estout- package to report the
> r2 and r2_a summary statistics for multiply-imputed data? If so, how?

I'll consider adding an option to -mibeta- to store the e(r2) and e(r2_a)
scalars in the future.  Meanwhile, the current version of mibeta.ado can be
modified to accommodate this.  The easiest way to do this is to add the
following line:

  mata: st_numscalar("e(`var')", st_numscalar("`m'"))

to the end of the _di_stats subprogram in mibeta.ado (between lines 146 and
147).  This will instruct -mibeta- to store the mean estimates of R2 and
adjusted R2 in -e(r2_1)- and -e(r2_2)-, respectively.  The location of the
current version of mibeta.ado can be found by typing:

  . which mibeta.ado

Nic's specification of the -stats()- option in -esttab- would then change to

  . esttab ..., ... stats(r2_1 r2_2, labels("R-Square" "Adjusted R-Square"))

I'll send Nic the modified mibeta.ado privately.

As a side note, you should be careful when using the -post- option with -mi
estimate: <command>-.  It does provide a way to access some of
<command>-specific postestimation tools but not all of them may be applicable
within multiple-imputation context.  This is one of the reasons why -mi
estimate- does not allow them by default.  Remember, with or without -post-,
the results are still multiple-imputation results.

For example, along with coefficients and standard errors Nic is also accessing
p-values.  P-values, as reported by -mi estimate-, are computed using MI
degrees of freedom, which are specific to each coefficient and are saved in
the -e(df_mi)- matrix.  Unless -esttab- has a special logic to accommodate
coefficient-specific degrees of freedom, the degrees of freedom and p-values
reported by -esttab- by default will not, in general, correspond to the
results from -mi estimate-.


-- Yulia
ymarchenko@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/