Drop parameters from omitted variables of e(b) and e(V)

[Alvaro Gutierrez]

Dear Statalist,

I have the following problem. I have written a command that uses ml to estimates a model, but I got my cluster/robust standard errors using a self-written routine in Mata that used e(V) and e(b) as input to do the computations (i.e., analytical gradients are written in Mata).

Everything works fine, but given this way to proceed, I am having problems to parse the e(V) and e(b)objectswhen there are omitted variables because of colinearity.

Hence, I would like to be able to:

• First, get only the covariates associated with the regressors that were not omitted from e(V) and e(b)objects.
• Second, using the objects from above, compute the robust/cluster variance-covariance matrices.
• Finally, I would like to put the values back in the original e(V) output keeping its original structure.

I only need help with the First point, but I illustrate the other two for the sake of concreteness, and also in case other, probably better, solutions could arise to solve my ultimate goal.

Below you might find a sketch that illustrates what I want (doing it mostly manually).


First, I run a regression with a repeated variable that is omitted. Here I have noticed that there is an o. that resembles the equation name of the matrices. Still, unfortunately, it doesn't follow their behavior, meaning I can't extract them with matrix omitted = e(b)[1,"o.:"] because I get the error equation o. not found.

clear all
sysuse auto , replace
version 12
tempvar constant  
gen `constant' = 1
reg price mp mp `constant' , nocons
matrix V = e(V)
mat li V

symmetric V[3,3]
                               o.            
                 mpg         mpg    __000000
     mpg   2817.1347
   o.mpg           0           0
__000000  -59997.356           0   1370802.5

This is what I would like to do automatically, keep only the parameters associated with variables that were not omitted.

**Doing manually what I want to automatize 
mat V_non_omitted = (V[1,1],V[1,3]) \ (V[1,3],V[3,3])
mat colnames V_non_omitted = "mpg" "`constant'" 
mat rownames V_non_omitted = "mpg" "`constant'" 
mat li V_non_omitted

symmetric V_non_omitted[2,2]
                 mpg    __000000
     mpg   2817.1347
__000000  -59997.356   1370802.5

This is an okay input (plus the same to e(b)) to my Mata procedure to compute robust/cluster standard errors. Therefore, just let's assume that I got my robust matrix V_robust.

*This should emulate my procedure to compute robust/cluster standard errors
mat V_robust = V_non_omitted *1.5
mat li V_robust 
*Robust Matrix
symmetric V_robust[2,2]
                 mpg    __000000
     mpg   4225.7021
__000000  -89996.033   2056203.8

Finally, I want to place back the robust errors in the same positions as they should be placed, and in that way, when invoking ml display, everything shows on-screen nicely. For now, the way that I am doing it is recovering the column and row number using colnum and rownum, respectively.

*Put back the std error where they belong 
foreach i in `row_names' {
    foreach j in `col_names' {    
        *recovers rownum  of variable `i'
        local row_var_i = rownumb(V,"`i'")
        *recovers colnum  of variable `j'
        local col_var_j = colnumb(V,"`j'")    
        *replace the robust values in the original var-covar.
        matrix V[`row_var_i', `col_var_j'] = V_robust["`i'", "`j'"]
}
}

symmetric V_robust[2,2]
                 mpg    __000000
     mpg   4225.7021
__000000  -89996.033   2056203.8

. mat li V

symmetric V[3,3]
                               o.            
                 mpg         mpg    __000000
     mpg   4225.7021
   o.mpg           0           0
__000000  -89996.033           0   2056203.8

Thank you in advance,

Álvaro

---

Stata 16.1 MP
Win10/Linux Mint 19.1
https://alvarogutyerrez.github.io/

[Alvaro Gutierrez]

A solution for the First point can be found in a cross-posted question on Stackoverflow.


Álvaro

---

Stata 16.1 MP
Win10/Linux Mint 19.1
https://alvarogutyerrez.github.io/

[Martin Andresen]

Doesn't -set emptycells drop- before the regression command fix this?