Bootstrap Error r(2000) "insufficient observations to compute bootstrap standard errors" in DFL Decomposition Analysis

[Clara Rummel]

Hi everyone,

I am trying to calculate the significances after applying a DFL detailed decomposition, where I have re-weighted all observations, using bootstrap. Unfortunately, the approaches suggested here in the forum have not worked so far, so I would really appreciate any help as I have been stuck on this problem for weeks.

. noisily bootstrap r(p101), reps(1000) seed(123456): dfl_dd
(running dfld on estimation sample)

Bootstrap replications (1000)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx    50
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   100
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   150
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   200
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   250
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   300
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   350
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   400
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   450
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   500
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   550
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   600
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   650
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   700
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   750
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   800
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   850
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   900
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx   950
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx  1000
insufficient observations to compute bootstrap standard errors
no results will be saved
r(2000);

Part of the program I am running:

program define dfl_dd, rclass
probit c_dummy $agegr_dummies $immigr_dummies $educ_dummies $lfs_dummies $incomeq_dummies [pw=hwgt], vce(robust)
predict double phat
gen double weight001new=(1-phat)/(phat)
gen weight001=.
replace weight001=weight001new
drop phat
// Second probit regression, excluding one set of variables
probit c_dummy $immigr_dummies $educ_dummies $lfs_dummies $incomeq_dummies [pw=hwgt],vce(robust)
predict double phat
gen double weight001hnew=(phat)/(1-phat)
replace weight001h=weight001hnew*weight001
gen double weight001h2=(1-phat)/(phat)
drop phat
*continue for the other 3 probit, always excluding a variable

In this part of the code, I am iterating over 120 different different pre-programmed reweighting schemes to calculate several summary statistics (mean, p10, p25, p50, p75, p90, p95) for the wealth variable. c_dummy is a country, as we are working on the wealth distribution in two countries.

*(...)
forvalues i=116(1)120{   
quietly su anw_ppp if c_dummy==1 [aw=x`i'`z'],detail   
foreach stat in mean p10 p25 p50 p75 p90 p95{   
local `stat'm`i' = r(`stat')
if `i'==116 local `stat'm1=r(`stat')    
if `i'==117 local `stat'm2=r(`stat')   
if `i'==118 local `stat'm3=r(`stat')   
if `i'==119 local `stat'm4=r(`stat')   
if `i'==120 local `stat'm5=r(`stat')   
}   
}   
foreach stat in mean p10 p25 p50 p75 p90 p95{   
local `stat'1`z'=``stat'1`z''+(``stat'm00'-``stat'm1')   
local `stat'2`z'=``stat'2`z''+(``stat'm2'-``stat'm3')   
local `stat'3`z'=``stat'3`z''+(``stat'm4'-``stat'm5')   
local `stat'4`z'=``stat'4`z''+(``stat'm3'-``stat'm4')   
local `stat'5`z'=``stat'5`z''+(``stat'm1'-``stat'm2')   
}   
}   
foreach stat in mean p10 p25 p50 p75 p90 p95{   
local `stat'1=(``stat'1a'+``stat'5b'+``stat'4c'+``stat'3d'+``stat'5e')/120   
local `stat'2=(``stat'2a'+``stat'1b'+``stat'5c'+``stat'4d'+``stat'4e')/120   
local `stat'3=(``stat'3a'+``stat'2b'+``stat'1c'+``stat'5d'+``stat'3e')/120   
local `stat'4=(``stat'4a'+``stat'3b'+``stat'2c'+``stat'1d'+``stat'2e')/120   
local `stat'5=(``stat'5a'+``stat'4b'+``stat'3c'+``stat'2d'+``stat'1e')/120   
local `stat'6=``stat'1'+``stat'2'+``stat'3'+``stat'4'+``stat'5'   *compositional effect
local `stat'7=``stat'm00'-``stat'c0'   *unadjusted difference

return scalar `stat'1=``stat'1'   
return scalar `stat'2=``stat'2'   
*continue for 3-7 
}  

*** dropping weight00, wgt and x1a-120e ***
*(...)
end

After this section of code follows the bootstrap code. I have already tried debugging with preserve etc, as recommended in other posts. My dummy variables of the five independent variables have at least 2000 observations in each category. If I run the programme dfl_dd without bootstrap, I can display the percentiles.

I hope I have been able to explain my problem and have shared all the important information, if not please excuse me, I am still a very newbie in the world of econometrics (btw, running a DFL right away when using Stata for the first time may not have been the smartest idea, but sweat and tears got me here and everything except for the bootstrap works :D)

Have a great day,
Clara Rummel