subsample comparison analysis

[Helen Chang]

Hi all, I am currently using the following code to compare the statistical difference for the coefficient between two subsamples.

qui reg DepVar IndVar if LOSS==1
est store m1

qui reg DepVar IndVar if LOSS==0
est store m2

suest m1 m2, cluster(firm)
test [m1_mean]IndVar-[m2_mean]IndVar=0


Is there a way that I can include both fixed effects (year fixed and industry fixed effects) and firm cluster in this setting? Since in other full sample tests, I use reghdfe DepVar IndVar, absorb (industry year) cluster(firm)

[Andrew Musau]

reghdfe is from SSC, as you are asked to explain in FAQ Advice #12. You can still do the test using this command, there is no need for suest here.

webuse grunfeld, clear
*SAMPLE SPLIT
gen group= _n<=_N/2
*SEPARATE REGRESSIONS
reghdfe invest mvalue kstock if !group, a(company year) cluster(company)
reghdfe invest mvalue kstock if group, a(company year) cluster(company)
*JOINT REGRESSION
reghdfe invest i.group#(c.mvalue c.kstock), a(i.group#company i.group#year) cluster(company)
*TEST MVALUE IN SAMPLE 1= MVALUE IN SAMPLE 2
test 0.group#c.mvalue= 1.group#c.mvalue

Res.:

. *SEPARATE REGRESSIONS

.
. reghdfe invest mvalue kstock if !group, a(company year) cluster(company)
(MWFE estimator converged in 2 iterations)

HDFE Linear regression                            Number of obs   =        100
Absorbing 2 HDFE groups                           F(   2,      4) =      28.21
Statistics robust to heteroskedasticity           Prob > F        =     0.0044
                                                  R-squared       =     0.9087
                                                  Adj R-squared   =     0.8762
                                                  Within R-sq.    =     0.5632
Number of clusters (company) =          5         Root MSE        =     9.6084

                                (Std. Err. adjusted for 5 clusters in company)
------------------------------------------------------------------------------
             |               Robust
      invest |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      mvalue |   .0814857   .0329346     2.47   0.069    -.0099553    .1729268
      kstock |   .1033723   .0158188     6.53   0.003     .0594523    .1472922
       _cons |  -5.356396   10.67978    -0.50   0.642    -35.00821    24.29541
------------------------------------------------------------------------------

Absorbed degrees of freedom:
-----------------------------------------------------+
 Absorbed FE | Categories  - Redundant  = Num. Coefs |
-------------+---------------------------------------|
     company |         5           5           0    *|
        year |        20           0          20     |
-----------------------------------------------------+
* = FE nested within cluster; treated as redundant for DoF computation

.
. reghdfe invest mvalue kstock if group, a(company year) cluster(company)
(MWFE estimator converged in 2 iterations)

HDFE Linear regression                            Number of obs   =        100
Absorbing 2 HDFE groups                           F(   2,      4) =      61.02
Statistics robust to heteroskedasticity           Prob > F        =     0.0010
                                                  R-squared       =     0.9416
                                                  Adj R-squared   =     0.9208
                                                  Within R-sq.    =     0.6795
Number of clusters (company) =          5         Root MSE        =    74.6136

                                (Std. Err. adjusted for 5 clusters in company)
------------------------------------------------------------------------------
             |               Robust
      invest |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      mvalue |   .1264177   .0147231     8.59   0.001     .0855399    .1672954
      kstock |   .3596463   .0811994     4.43   0.011     .1342006     .585092
       _cons |  -118.5172    35.7848    -3.31   0.030    -217.8717   -19.16265
------------------------------------------------------------------------------

Absorbed degrees of freedom:
-----------------------------------------------------+
 Absorbed FE | Categories  - Redundant  = Num. Coefs |
-------------+---------------------------------------|
     company |         5           5           0    *|
        year |        20           0          20     |
-----------------------------------------------------+
* = FE nested within cluster; treated as redundant for DoF computation

.
. *JOINT REGRESSION

.
. reghdfe invest i.group#(c.mvalue c.kstock), a(i.group#company i.group#year) cluster(company)
(MWFE estimator converged in 2 iterations)

HDFE Linear regression                            Number of obs   =        200
Absorbing 2 HDFE groups                           F(   4,      9) =      49.94
Statistics robust to heteroskedasticity           Prob > F        =     0.0000
                                                  R-squared       =     0.9559
                                                  Adj R-squared   =     0.9398
                                                  Within R-sq.    =     0.6781
Number of clusters (company) =         10         Root MSE        =    53.1955

                                 (Std. Err. adjusted for 10 clusters in company)
--------------------------------------------------------------------------------
               |               Robust
        invest |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
---------------+----------------------------------------------------------------
group#c.mvalue |
            0  |   .0814857   .0311293     2.62   0.028     .0110663    .1519052
            1  |   .1264177    .013916     9.08   0.000     .0949374    .1578979
               |
group#c.kstock |
            0  |   .1033723   .0149517     6.91   0.000     .0695492    .1371954
            1  |   .3596463   .0767486     4.69   0.001     .1860288    .5332637
               |
         _cons |  -61.93679   17.64875    -3.51   0.007     -101.861   -22.01254
--------------------------------------------------------------------------------

Absorbed degrees of freedom:
---------------------------------------------------------+
     Absorbed FE | Categories  - Redundant  = Num. Coefs |
-----------------+---------------------------------------|
   group#company |        10          10           0    *|
      group#year |        40           0          40     |
---------------------------------------------------------+
* = FE nested within cluster; treated as redundant for DoF computation

.
. *TEST MVALUE IN SAMPLE 1= MVALUE IN SAMPLE 2

.
. test 0.group#c.mvalue= 1.group#c.mvalue

 ( 1)  0b.group#c.mvalue - 1.group#c.mvalue = 0

       F(  1,     9) =    1.74
            Prob > F =    0.2202

[marie nour]

Dear Andrew Musau

I am running a similar regression where my groups are made of south-south, north-south and north-north countries. And i run a regression like this:

 
 reghdfe lnimports lndist comlan border if group==3, a(importer exporter) cluster(pair)

Then, I create 3 dummies (NN, NS and SS) and their interaction terms with each independent variable and run the following regression:

 
 reghdfe lnimports lndist comlan border NNlndist NNcomlan NNborder NSlndist NScomlan NSborder, a(importer exporter) cluster(pair)

I am wondering what happens with the FE if I leave them like that (instead of interacting them)? In the theory, country FE account for multilateral resistance terms (Anderson and van Wincoop, 2004), so they should be computed relative to all countries (and not to the subsample). So is this the way I did it the right way to go ?
Thank you

[Andrew Musau]

reghdfe lnimports lndist comlan border NNlndist NNcomlan NNborder NSlndist NScomlan NSborder, a(importer exporter) cluster(pair) [/CODE]

From a fixed effects perspective, that is fine as long as the dummies vary over importers and exporters. In this case, you just have dummies representing country groupings included in the fixed effects model. On the correctness of the specification, you need to consult with an expert on gravity models, but there is no technical issue.