Fastest way to collect values of regression coefs

[Park Sungho]

I want the values of only 160 of the god knows how many coefficients in my regression.

I am using reghdfejl but things are not much faster. Is there any stata command or method where I can simply obtain the values of the coefficients for the variables I want? I do not even need the SEs as I will bootstrap them. (this is for a program)

[Carlo Lazzaro]

Park:
welcome to this forum.
Did you take a look at -statsby-?

[Park Sungho]

Park:
welcome to this forum.
Did you take a look at -statsby-?

I don't think this would preserve the point estimates. You see, I am estimating this model with a bunch of fixed effects and controls....

reghdfejl y [coefs of interst varying by town] control vector, absorb(lots of FEs), cluster(id)

is what I am doing

[Julian Reif]

Have you tried regsave?

sysuse auto, clear
reghdfe price mpg displacement, absorb(foreign)
regsave
list

     +----------------------------------------------------+
     |          var       coef     stderr    N         r2 |
     |----------------------------------------------------|
  1. |          mpg   -98.8856   63.17063   74   .4631095 |
  2. | displacement   22.40416   4.634239   74   .4631095 |
  3. |        _cons   3850.973   2070.066   74   .4631095 |
     +----------------------------------------------------+

[FernandoRios]

Hi Park
if you are "absorbing" too many fixed effects, that will most likely increase the number of iterations, thus time needed, to estimate all required coefficients.
for example 1 fixed effects requires only 1 pass to demean the data
2 sets would require (in my experience) about 10-40 passes before the demean is ready (within margins of errors)
with 3 or more, this would increase even more.

So, better alternative. Choose which FE have few enough cases, and add them as controls, rather than fixed effects.
Or rethink about your model specificaiton
F

[Park Sungho]

Hi Park
if you are "absorbing" too many fixed effects, that will most likely increase the number of iterations, thus time needed, to estimate all required coefficients.
for example 1 fixed effects requires only 1 pass to demean the data
2 sets would require (in my experience) about 10-40 passes before the demean is ready (within margins of errors)
with 3 or more, this would increase even more.

So, better alternative. Choose which FE have few enough cases, and add them as controls, rather than fixed effects.
Or rethink about your model specificaiton
F

I run this routine, and all processes exit with errors

cap program drop investment_rat3
program define investment_rat3, eclass
global main_restrictions slead1_k1<50 & i_rate<10 & rel_year>=-5 & rel_year<5
global main_restrictions2 slead1_w<10 & rel_year>=-5 & rel_year<5
global main_restrictions3a pgrowth>-2.621353 & pgrowth<2.114623 & rel_year>=-5 & rel_year<5
global fixed_effects i t i.t#c.praz ///
i.sector2#c.GDP_growth i.score#c.size_pctile1 ///
i.t#c.GDP_growth
****************************************************************************
* 2c. First regression to obtain lambdas
****************************************************************************
preserve
reghdfejl cas pc_trib_N*  STUFF, absorb(STUFF) compact
restore
preserve
****************************************************************************
* 2d. Calculate lambda_pre and lambda_post based on regression
****************************************************************************
gen lambda_pre = .
gen lambda_post = .
forval i = 1/160 {
cap scalar lambda_`i' = _b[pc_trib_N`i']
replace lambda_pre = lambda_`i' if trib_N_pre == `i'
replace lambda_post = lambda_`i' if trib_N_post == `i'
}
gen lambda_diff = lambda_post - lambda_pre
gen post_treatment = lambda_diff * (rel_year>=0) if dataset1==0
cap drop lambda_*
****************************************************************************
* 2e. Final regression of interest
****************************************************************************
reghdfejl STUFF post_treatment slead1_k1 if $main_restrictions & dataset1==0, ///
absorb($fixed_effects) vce(cl codice_f)
/*
Post the coefficient vector into e(b). We also post the estimation
sample for bootstrap to know which observations were used.
*/
matrix b = e(b)
// 2. Post them as e(b) and e(V)
ereturn post b
restore
end