Wooldridge(2009) How to use stata to calculate TFP by this literature.

[Qing Tong]

Recently, I have read this literature. As result of this mean of calculating TFP is different. I want to know how to complete by stata.
On estimating firm-level production functions using proxy variables to control for unobservables This is the name of the literature. I have tried many times to upload it. But not success

[Phil Bromiley]

You're assuming we either know Wooldridge so well we know what TFP means and how he wants to calculate TFP, or we're willing to go to the trouble of looking it up to help you. TFP is not in the index of my 2002 Wooldridge.

Look at the FAQ on asking questions. If you can tell us clearly without undefined acronyms what you're trying to estimate, we might be better able to help.

[Gabriele Rovigatti]

Hello everyone,

I agree with Phil that the question is not clear, at best. However, I presume that "TFP" in this context refers to Total Factor Productivity and that Qing was interested in the estimation methods Wooldrisdge presented in its 2009 paper "On estimating firm-level production functions using proxy variables to control for unobservables".

You may want to give prodest a try

ssc install prodest

It is a new Stata module aimed at estimating production functions using exaclty those models - please see the

help prodest

for clickable examples and explanations on program usage.

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

[Charlie Joyez]

Nice program Gabriele !
I've been using levpet (Levinsohn - Petrin) until now, but it is nice to have a single program to run alternative TFP computations.
Thanks for sharing it here.

[Gabriele Rovigatti]

Dear Charlie,

thanks for your post. Levpet has been one of our references - probably the most important - while writing the command. Hence, you should find the syntax familiar.

I look forward to receiving any suggestion/comment you may have.

Best,

Gabriele

[Salem Abdussalam]

Many thanks dear Gabriele for your suggestion on "ssc install prodest". It provided me with a new dimension to compare my estimations using different approaches.

With Kind Regards.

Salem.

[Rodrigo Remedio]

The prodest is a great contribution, Gabriele Rovigatti. I saw on the prodest code you use the Stata's gmm module, which seems to me to be very consistent with the Wooldridge proposition. I say that, because there are popular code snippets used by some authors that applies the Wooldridge methodology through the ivreg2/ivregress commands. They use lagged polynomials of the proxy and state variables as regressors, and lagged free variables as instrument for the free variable. This is done by Levisoh and Petrin here, and by Galuscak and Lizal here and here.
Anyway,would it be too hard to compare (theoretically) this with your implementation?

[Gabriele Rovigatti]

Dear Rodrigo,

thank you very much for your comment. I agree with you that the system gmm implementation is more aligned with the letter of Wooldridge (2009) and it is more clear and readable in terms of code. At the end of the day, that's the reason why we chose to implement it that way.

By the way, I think that the ivreg2 implementation has some limitations with respect to ours (even if it is extremely faster). Enclose you can find a couple of tables referring to both prodest and ivreg2. You will see that estimates are very similar but I have run the same commands on a number of datasets and results differ, sometimes dramatically. In a word, I suggest you to be extremely careful in using the iverg2 implementation and to proceed by comparison with similar models (LP/gmm wooldridge/ACF) when working with it.

ivreg2:

2-Step GMM estimation
---------------------

Estimates efficient for homoskedasticity only
Statistics consistent for homoskedasticity only

                                                      Number of obs =     1284
                                                      F(  8,  1275) =   497.52
                                                      Prob > F      =   0.0000
Total (centered) SS     =  3194.712607                Centered R2   =   0.7587
Total (uncentered) SS   =  234252.8329                Uncentered R2 =   0.9967
Residual SS             =  770.8340891                Root MSE      =    .7748

------------------------------------------------------------------------------
       log_y |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
    log_lab1 |   .1665038   .0298481     5.58   0.000     .1080026     .225005
    log_lab2 |    .083068   .0219718     3.78   0.000     .0400041    .1261319
       log_k |   .2166239    .051149     4.24   0.000     .1163738    .3168741

prodest:

prodest log_y, free(log_lab1 log_lab2) state(log_k) proxy(log_materials) va met(wrdg) poly(2) id(id) t(year)

wrdg productivity estimator

Dependent variable: value added                 Number of obs      =      1284
Group variable (id): id                         Number of groups   =       386
Time variable (t): year
                                                Obs per group: min =         1
                                                               avg =       4.6
                                                               max =        12

------------------------------------------------------------------------------
       log_y |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
log_lab1     |
    log_lab1 |   .1631629   .0234218     6.97   0.000     .1172569    .2090689
-------------+----------------------------------------------------------------
log_lab2     |
    log_lab2 |   .0955509   .0190286     5.02   0.000     .0582555    .1328463
-------------+----------------------------------------------------------------
log_k        |
       log_k |    .214541   .0594993     3.61   0.000     .0979245    .3311575
------------------------------------------------------------------------------

As for the theoretical comparison, at the moment I don't have any reference for that. I will come back to you as soon as I'll find any.

Best,

Gabriele

[Gabriele Rovigatti]

Dear all,

I will bump the thread for those interested in prodest. A new, more complete version of the command is now on the SSC and I complemented it with a working paper describing the whole suite, options and some comparative results downloadable here.

Hence, I strongly recommend to ssc install prodest, replace your systems and, for those who happen to use prodest in their papers, please cite:

Mollisi, Vincenzo and Rovigatti, Gabriele, 2017. "Theory and Practice of TFP Estimation: The Control Function Approach Using Stata", CEIS Working Paper No. 399.

Best,

Gabriele

[Ibnu Nur Hamzah]

Dear Gabriele,

Thank you so much for the package. I am using Woolridge (2009) for estimating TFP and I found prodest package is very helpful.
I have a question to ask for control(varlist) option within the module, are they control variable for TFP (i.e firm size, firm age, market concentration, etc)?

Thank you,

Kind Regards,
Ibnu

[Gabriele Rovigatti]

Dear Ibnu,

thanks for your post. I am not sure to understand what you mean with "control variable for TFP": you can think of the control(varlist) option in prodest as you would do in any other regression framework. In particular, in Wooldridge (2009)'s terms the control variables enter equations (3.8) as an additional term: you'll have

r_{it1}(\theta) = y_{it} - \alpha_{0} - w_{it}\beta - x_{it}\gamma - c_{it}\lambda - controls_{it} \zeta
r_{it2}(\theta) = y_{it} - \eta_{0} - w_{it}\beta - x_{it}\gamma - \rho_{1}(c_{i,t-1} \lambda) - ... - \rho_{G}(c_{i,t-1} \lambda)^{G} - controls_{it} \zeta

I think that, provided that this is consistent with your setting, you can use whatever kind of control.

I hope to have clarified.

Best,

Gabriele

[Ibnu Nur Hamzah]

Dear Ibnu,

thanks for your post. I am not sure to understand what you mean with "control variable for TFP": you can think of the control(varlist) option in prodest as you would do in any other regression framework. In particular, in Wooldridge (2009)'s terms the control variables enter equations (3.8) as an additional term: you'll have

r_{it1}(\theta) = y_{it} - \alpha_{0} - w_{it}\beta - x_{it}\gamma - c_{it}\lambda - controls_{it} \zeta
r_{it2}(\theta) = y_{it} - \eta_{0} - w_{it}\beta - x_{it}\gamma - \rho_{1}(c_{i,t-1} \lambda) - ... - \rho_{G}(c_{i,t-1} \lambda)^{G} - controls_{it} \zeta

I think that, provided that this is consistent with your setting, you can use whatever kind of control.

I hope to have clarified.

Best,

Gabriele

I see.... thank you so much for your answer. I am working on it now.

Cheers

[Mislav Sagovac]

Maybe, it is a stupid question, but have can I measure productivity for every firm after I have estimated production function using prodest?

I would need productivity (tfp) for every observation in panel.

[Gabriele Rovigatti]

Dear Mislav,

I'd suggest you to use the prodest predict postestimation program, provided with the command itself. Look at the helpfile by typing

help prodest_p

and - in case of doubt - shoot me an email at [email protected]

Best,

Gabriele