Display or Show joint test after estimating REGHDFE

[Jerry Roc]

Hello guys,

I run the REGHDFE command to estimating this: reghdfe ly $quatre $cinq, a (cmap dveduc3 t). On my output don't have the "joint test" of the 3 fixed effects groups. How can I display the joint F-test for my 3 fixed effects (cmap dveduc3 t).

I use: display e(F) but this command return me the F (. , .) = . value that we see generally in the top of the output regression table. Can someone please helping me with that issue.


I attached an output example, to help you guys understanding what I want to display (So the forth line of the table). The output example can from the internet.

Thank you!

[Andrew Musau]

reghdfe by Sergio Correia is from SSC. To access estimation results, run

ereturn list

The statistic is stored as

 e(F_absorb)

By the way, you are running an old version of the program. The current version does not display this test as part of the default output.

[Jerry Roc]

Hello, Thank you so much for your answer.

By the way, when I run : e(F_absorb), Stata gives me this error.

[Andrew Musau]

display e(F_absorb)

following the regression.

[Jerry Roc]

There is nothing that appears after doing this code.

[Andrew Musau]

You cannot run that line independent of the regression. In the do-file, run the entire block of code starting from the reghdfe command. Note that I invoke the -old- option in my regression command as I have the latest version of reghdfe.

sysuse auto
reghdfe mpg price weight, a(turn) old
di e(F_absorb)

. sysuse auto
(1978 Automobile Data)

. reghdfe mpg price weight, a(turn) old
(running historical version of reghdfe)
(dropped 4 singleton observations)
(converged in 1 iterations)

HDFE Linear regression                            Number of obs   =         70
Absorbing 1 HDFE group                            F(   2,     54) =      12.76
                                                  Prob > F        =     0.0000
                                                  R-squared       =     0.6980
                                                  Adj R-squared   =     0.6141
                                                  Within R-sq.    =     0.3210
                                                  Root MSE        =     3.4819

------------------------------------------------------------------------------
         mpg |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       price |  -.0001178   .0002235    -0.53   0.600    -.0005659    .0003304
      weight |  -.0055865    .001645    -3.40   0.001    -.0088846   -.0022885
-------------+----------------------------------------------------------------
    Absorbed |         F(13, 54) =      0.772   0.693             (Joint test)
------------------------------------------------------------------------------

Absorbed degrees of freedom:
---------------------------------------------------------------+
 Absorbed FE |  Num. Coefs.  =   Categories  -   Redundant     |
-------------+-------------------------------------------------|
        turn |           14              14              0     |
---------------------------------------------------------------+

. di e(F_absorb)
.7724645

[Jerry Roc]

Thank so much!!!!!

You had to put "old" after the code to be able to have it. Like : reghdfe y x, a(z) old

Thank you.

[Jerry Roc]

By the way, I am seeking to regress the wage gap between natives and recent immigrants according to the census year (t), the census metropolitan area (cmap) and the level of education (dveduc3); on a packet of predictors. My dependent variable is therefore the ratio of the average wage of natives to that of recent immigrants. To do this, from the observations of each census I created cells using the command collapse. For example, for the average salary of natives, I did: collapse w if n == 1, by (t cmap dveduc3). Where n is a dummy variable equals 1 if respondents are natives and 0 otherwise. The same logic applies to the average salary of recent immigrants as well as the explanatory variables of the model.

Since these are cells of different sizes because of the number of observation that varies in cmap for example. Sometimes cmap receive more migrants than others. So, I would like to use in weight in my regression the size of the cells (aw = count) where count is the number of observations in the cell. The idea is that if the residue has a variance sigma ^ 2 / n, the use of cells of different size implies that some cells havea variance of the error term lower than others because the averages have been estimated on more observations. Recognizing this and exploiting it allows efficiency gains. This is a generalized least squares estimator that is simply done by OLS using cell size as the weight. It should help the accuracy of my estimates.

How can I create the count which will take the place of the size of the cells?

[Andrew Musau]

Something like?

egen group= group(t cmap dveduc3)
bys t cmap dveduc3: egen freq= count(group) if !missing(w) & n == 1
collapse w freq if n == 1, by (t cmap dveduc3)

[Jerry Roc]

In fact, here's what I use as code:

reghdfe ly $quatre $cinq [aw=count], a (cmap dveduc3 t) old

The idea is, count refer to the cell size. I would like to know how create the count or the cell size to use it as a weight.

[Andrew Musau]

Doesn't the variable "freq" in #9 give you the count that you need? Otherwise I am not understanding what is needed. Here, freq gives you how many observations (i.e., combinations of t, cmap & dveduc3) were used to generate the average value of w. If the calculation is correct, you then use

[aw=freq]

If not, give a data example using dataex showing how the weight should look like. Stata 15, 16 (or a fully updated version of 14), read through

help dataex

For a previous version

ssc install dataex
help dataex

[Jerry Roc]

Thank you Mr Andrew Musau for your help.