-mi test- and specification of factor variable coefficients

[Stephen Jenkins]

help mi_test suggests that the specification of plain vanilla tests after mi estimate are of the form mi test coeflist .How does one specify coeflist in the case of factor variables? I can't get this to work (see below) for the case of a single factor variable, let alone interaction terms for 2 factor variables.

I can't find advice on this in the help or manuals (but may have missing something). I can get results if I specify the coeflist manually, but I want to avoid that.

 . mi estimate: reg gini_net  i.year if region == 5 & year > 1990 & year < 2007
  
Multiple-imputation estimates                     Imputations     =        100
Linear regression                                 Number of obs   =        401
                                                  Average RVI     =     0.0468
                                                  Largest FMI     =     0.0604
                                                  Complete DF     =        385
DF adjustment:   Small sample                     DF:     min     =     355.39
                                                          avg     =     363.17
                                                          max     =     378.04
Model F test:       Equal FMI                     F(  15,  382.8) =       0.53
Within VCE type:          OLS                     Prob > F        =     0.9249
  
------------------------------------------------------------------------------
    gini_net |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        year |
       1992  |   .2601348   1.525127     0.17   0.865    -2.738659    3.258929
       1993  |   1.006815     1.5231     0.66   0.509    -1.988172    4.001803
       1994  |   1.377235   1.527145     0.90   0.368    -1.625771     4.38024
       1995  |   1.623852   1.514133     1.07   0.284    -1.353781    4.601486
       1996  |   1.672173   1.519782     1.10   0.272    -1.316452    4.660798
       1997  |   1.816323   1.521787     1.19   0.233    -1.176278    4.808924
       1998  |   2.290752   1.540125     1.49   0.138    -.7379953    5.319499
       1999  |    1.96233   1.520791     1.29   0.198    -1.028517    4.953178
       2000  |   2.744941    1.56078     1.76   0.079    -.3245387     5.81442
       2001  |   2.897571   1.561122     1.86   0.064    -.1725886    5.967731
       2002  |   1.074202   1.546792     0.69   0.488    -1.967778    4.116182
       2003  |   2.202671   1.569541     1.40   0.161    -.8839185     5.28926
       2004  |   1.501412   1.560643     0.96   0.337    -1.567526    4.570351
       2005  |   1.584315   1.597405     0.99   0.322    -1.556956    4.725586
       2006  |    2.01109   1.657117     1.21   0.226    -1.247545    5.269725
             |
       _cons |   44.85401   1.128164    39.76   0.000     42.63529    47.07273
------------------------------------------------------------------------------
  
.
end of do-file
  . mi test i.(1991/2006)year
note: assuming equal fractions of missing information
i:  operator invalid
r(198);
  
. mi test 1991.year 1992.year 1993.year 1994.year 1995.year 1996.year 1997.year 1998.year 1999.year 2000.year 2001.year 2002.year 2003.year 2004.year 2005.year 2006.year
note: assuming equal fractions of missing information
   ( 1)  1991b.year = 0
 ( 2)  1992.year = 0
 ( 3)  1993.year = 0
 ( 4)  1994.year = 0
 ( 5)  1995.year = 0
 ( 6)  1996.year = 0
 ( 7)  1997.year = 0
 ( 8)  1998.year = 0
 ( 9)  1999.year = 0
 (10)  2000.year = 0
 (11)  2001.year = 0
 (12)  2002.year = 0
 (13)  2003.year = 0
 (14)  2004.year = 0
 (15)  2005.year = 0
 (16)  2006.year = 0
       Constraint 1 dropped
         F( 15, 382.8) =    0.53
            Prob > F =    0.9249

.

[daniel klein]

Stephen, I did not try this, but it seems you have a typo in

mi test i.(1991/2006)year

The dot has to be after the parentheses as in

mi test i(1991/2006).year

Best
Daniel

[Stephen Jenkins]

Thanks for looking into this, Daniel. Unfortunately, your suggestion leads to an error as well:

. mi test i(1991/2006).year
note: assuming equal fractions of missing information
i not found
r(111);

If I can beat this one, I guess we'll be able to beat the problem of dealing with interacted factor variables as well.

[Alfonso Sánchez-Peñalver]

I'm not sure it will work, but what if you leave the i out of the expression?

mi test (1991/2006).year

What made me think of this is that when you specify the test year by year, you don't include the i prefix before the years. As I said, not sure it will work but I figure it's worth a try.

[skolenik]

First off, none of these suggestions actually work for the plain test:

sysuse auto, clear
reg mpg i.rep78 foreign if _m==0
test rep78
test i.rep78
test i(2/5).rep78
test (2/5).rep78

That is, not until you use the testparm form (which is not supported by mi)

testparm i(2/5).rep78
testparm i.rep78

To turn this around and develop a working solution specifically in terms of test, avoiding cumbersome typing, consider

levelsof rep78 if e(sample), local(theselevels)
local rep78levels = subinstr("`theselevels' "," ",".rep78 ", .)
test `rep78levels'

Note there is a space after `theselevels' in the string manipulation command. It is needed to correctly append .rep78 to the last level encountered. Note also that the whole thing is fragile and subject to the limits of Stata on the length of strings and string expressions. With this in hand, we can try this on mi, as well:

mi set wide
mi register imputed rep78
mi register regular mpg foreign length
mi impute pmm rep78 = mpg foreign length, add(10)
mi estimate : reg mpg i.rep78 foreign
mi test `rep78levels'

[daniel klein]

As a canned work-around you may also use fvexpand, as in

sysuse auto
fvexpand i.rep78
ret li

Probably you want to exclude the base level.

However, I would like to see the a mi testparm command in the future, which I imagine should not be too hard to implement. You may write your own wrapper.

Best
Daniel

[Yulia Marchenko (StataCorp)]

As Stas pointed out, test and mi test do not support the varlist specification that Stephen desires but rather the coeflist specification; you can read more about the definition of coeflist in the help file for the test command. The varlist specification is supported only by testparm. We will consider adding an mi testparm command in the future.

Meanwhile, Stephen can use the _ms_extract_varlist command to convert the factor-variable specification into a list of "factor coefficients" as stored in e(b), and use that list with mi test or test. This is how testparm is handling the varlist specification. For multiple-equations commands, you may also use the equation() option to choose the appropriate equation; type help _ms_extract_varlist for details.

Below, I show several examples for test but they can be applied directly to mi test.

. sysuse auto, clear
. regress mpg weight i.rep78
<output omitted>

. _ms_extract_varlist i.rep78, noomitted
. test `r(varlist)'

 ( 1)  2.rep78 = 0
 ( 2)  3.rep78 = 0
 ( 3)  4.rep78 = 0
 ( 4)  5.rep78 = 0

       F(  4,    63) =    1.10
            Prob > F =    0.3668

. testparm i.rep78

 ( 1)  2.rep78 = 0
 ( 2)  3.rep78 = 0
 ( 3)  4.rep78 = 0
 ( 4)  5.rep78 = 0

       F(  4,    63) =    1.10
            Prob > F =    0.3668

We can refer to a range of levels of a factor variable.

. _ms_extract_varlist i(2/4).rep78, noomitted
. test `r(varlist)'

 ( 1)  2.rep78 = 0
 ( 2)  3.rep78 = 0
 ( 3)  4.rep78 = 0

       F(  3,    63) =    0.02
            Prob > F =    0.9964

. testparm i(2/4).rep78

 ( 1)  2.rep78 = 0
 ( 2)  3.rep78 = 0
 ( 3)  4.rep78 = 0

       F(  3,    63) =    0.02
            Prob > F =    0.9964

We can perform estimation using a subsample and automatically obtain the levels of the factor variable relevant to that subsample.

. regress mpg weight i.rep78 if foreign==1
<output omitted>

. _ms_extract_varlist i.rep78, noomitted
. test `r(varlist)'

 ( 1)  4.rep78 = 0
 ( 2)  5.rep78 = 0

       F(  2,    17) =    4.00
            Prob > F =    0.0378

. testparm i.rep78

 ( 1)  4.rep78 = 0
 ( 2)  5.rep78 = 0

       F(  2,    17) =    4.00
            Prob > F =    0.0378

[Stephen Jenkins]

Thank you for all the comments. And Yulia: thanks too for being very diplomatic about my lack of the appreciation of the distinction between the varlist specification and the coeflist specification -- and of course for the info about _ms_extract_varlist.