Anyone knows about measuring Total Factor Productivity (TFP) by ACF(Ackerberg, Caves & Frazer, 2006) method?

[Mohiuddin Alamgir]

Dear All,

I have been measuring the TFP by ACF method and for that, I have taken

Jagadeesh Shivadashan, Asssociate Professor, University of Michigan,

code:[http://webuser.bus.umich.edu/jagadee.../acf_code.html]. It helped me measuring the coeficients of Labors and Capitals after 100 iterations. However, all the coefficients are same in all the iterations. Now, my problem is, I just got the coefficients, now have to find the residuals by which is actually TFP by "levpet" or "opreg" command, may be. I only have the rawmaterials values, i.e. I have to do it by "levpet" command. But, I am now helpless since, in the Jagadeesh Shivadashan's code, there is nothing that tells me how to get the TFP.

For your convenient I am giving the log files here, and attaching the final output coefficients that have I found from these codes. I would really appreciate if someone comes up with the solutions.
The Jagadeesh Shivadashan's (with all the respect to him) code is following while I just customized accordingly.
Here, lnva is log value added, lnb is log (blue collar employment), lnw is log(white collar employment) and lnm is log(real material cost)

////** This code generates 100 block-boostrapped samples of the data, and then calls
///the acf_solution.do program to generate corresponding ACF production function coefficient
///estimates.  The directory/data file names  need to be modified as necessary ***/

use "VA_L_11_12_K_1.dta", clear

/** This counter sets the number of bootstrap iterations you want to have*/
global nrep 100
matrix A =J($nrep,4,99)
*****************************
///** Here we start the code to allow us to bootstrap the sample **/
sort panelid lyear
//egen gid=group(panelid)
//qui su gid
scalar tt=r(max)
//sort gid year
by panelid: gen count=_N
by panelid: gen smalln=_n
sort panelid smalln
save tempk_me, replace
local i=1
while `i' < $nrep + 1 {
use tempk_me, clear
///** Here we draw random samples with replacement. We sample each
 ////   set of plant level observations as an independent block.**/
/**quietly {
keep if gid[_n]~=gid[_n-1]
sort gid
keep gid count
save temp1_me, replace
set seed `i'
gen kk=int(tt*uniform()+1)
keep kk
gen test=99
rename kk gid
sort gid
merge n:1 gid using temp1_me
drop _merge
keep if test~=.
drop test
egen fill=fill(1 2/3)
expand count
sort fill gid
by fill: gen smalln=_n
sort gid smalln
merge n:1 gid smalln using tempk_me
drop _merge
drop if fill==.
ren panelid  oldpanelid
ren fill panelid
save temp2_me, replace
}**/


*********************************
disp "Bootstrap replication `i'"
do for_ACF.do

matrix A[`i',1]=`i'
matrix A[`i',2]=vcoef[1,1]
matrix A[`i',3]=vcoef[1,2]
matrix A[`i',4]=vcoef[1,3]
local i=`i'+1
}

matrix coln A = iter_num bcoef wcoef kcoef
svmat A, name(col)

keep iter_num *coef
keep if iter_num~=.

foreach xx in b w k {
qui su `xx'coef, d
scalar `xx'mean=r(mean)
scalar `xx'99=r(p99)
scalar `xx'1=r(p1)
}
display "bmean" bmean " b99" b99 " b1" b1 " wmean" wmean " w99" w99 " w1" w1  " kmean" kmean " k99" k99 " k1" k1


/** The 1% and 99% gives the bootstrap confidence interval ***/

and the "do for_ACF.do" The codes are:

/** This program generates estimates using one version of the ACF moment conditions **/
set more off
/**  Ackerberg-Caves-Frazer first stage estimation equation:  getting to phihat**/
/*  Generating 2nd order proxy polynomial  */
local i=1
foreach x in lnb lnw lnk lnm {
gen double var_`i'=`x'
local i=`i'+1
}

forv i=1/4{
forv j=`i'/4{
gen double var_`i'`j'=var_`i'*var_`j'
}
}

/** In ACF, no coefficients are identified in the first stage. phihat is simply the predicted y*/
qui regress lnva var*
qui predict phihat
xtset panelid lyear
/** Once we have this phihat, we need to enter a search routine to find the best estimates for the capital
and labor coefficients.  The objective function we want to maximize is defined below --
the  moment conditions assume current values of capital and labor are orthogonal to
unpredicted parts of the productiviy term, based on the asumption that all of labor and capital are chosen in period
t-1 ***/

capture program drop ofn
program define ofn
tempvar e omega omlag omlag2 omlag3 epsilon mom1 mom2 mom3 xx
matrix score `e'=`1'
qui gen double `omega'=phihat -`e'
sort panelid lyear
qui gen double `omlag'=L.`omega'
qui gen double `omlag2'=`omlag'^2
qui gen double `omlag3'=`omlag'^3

qui reg `omega' `omlag' `omlag2' `omlag3'
qui predict `xx'
qui gen `epsilon'=`omega' -`xx'

    qui gen double `mom1'= (`epsilon'*lnk)
    qui su `mom1'
    scalar sumom1= (r(sum))^2

    qui gen double `mom2'= (`epsilon'*lnb)
    qui su `mom2'
    scalar sumom2= (r(sum))^2

    qui gen double `mom3'= (`epsilon'*lnw)
    qui su `mom3'
    scalar sumom3= (r(sum))^2
    scalar obj= -(sumom1 + sumom2 +sumom3)
    scalar `2'=obj
end

/** Using the amoeba optimization routine ***/
***********************************************************
qui reg lnva lnb lnw lnk
mat initols=e(b)
amoeba ofn initols obj vcoef . 200 0.0000001

The output file is attached here:

[Hongtao Liu]

You would need to download and install amoeba.ado ? how to do it?

[Gabriele Rovigatti]

Hi everyone,

you may want to give prodest a try. To my knowledge, it is the only Stata module currently able to estimate ACF.
Code:

ssc install prodest

https://ideas.repec.org/c/boc/bocode/s458239.html

https://papers.ssrn.com/sol3/papers....act_id=2916753

Please note that it is still a beta version: every suggestion, comment or bug reporting will be more than welcome. Please report any issue you may have to: [email protected], or write here.

Good luck,

Gabriele