[Atsuyuki Kato]

Dear Gabriele

Thank you for your help and suggestion.

You are right.
The program was stuck because my user account includes a space.
So, I added another local account that includes no space.
On this account, I successfully got the results.

Best

Atsuyuki Kato

[Mark Pham]

Hi Gabriele,

I have used -prodest- for one of my project and I found the command extremely handy and customizable. I encounter the following issues while running the code and was wondering if you could help explain what is going on.

When I use the attrition option, the output shows that "No ID exits the sample. Running the estimation with no attrition". The problem are that:
1. it is obvious in my panel that firms exit
2. This warning appears at different stages of bootstrap replications for different methods (even though with the same data)
My guess is that this warning depends on the method I use, but it's not intuitive to me why and how would this affects my results.

The other issue I have encountered is that in one of the estimation using -lp- with -acf- correction, the standard errors are missing even though it seems that I have enough replication. The output is captured in attached photo.


Thank you and I really to hear from you,
Sincerely,
Mark

[Gabriele Rovigatti]

Dear Mark,

thanks for the message.

I will try to address the issues you raised below:

1) the "No ID exits the sample. Running the estimation with no attrition" warning message, as the content suggests, it is triggered whenever no firm appears to exit the sample. In your case, the original sample does not have this issue, and in fact the warning message appears only at the 21st bootstrap repetition; as you know, in fact, prodest implements a series of cluster bootstrap repetitions, hence proceeds in reshaping the original data at the id level. Sometimes, this might cause all exiting firms not to be in the sample, and that's where the warning message comes from. Let me stress, though, that it would not affect your estimates nor it will bias them (indeed, thanks for pointing this unexpected behaviour: I will fix it in the next issue of the command).

2) as I tried to point out in my paper, ACF-corrected models are particularly affected by the starting points of the optimization routine for the second stage. In prodest, they are taken by default by the first stage (that is:\hat{\beta} + noise); however, sometimes (it depends on the shape of the GMM problem) the second stage is stuck in local maximum at the starting points. As a result, the parameter estimates do not move, and the standard errors cannot be computed. I would suggest to use the undocumented option init() to overcome the issue: in particular, you may want to try to run

prodest lnrva, free(lnl) proxy(lnrm) state(lnk) poly(3) met(lp) valueadded acf attrition reps(50) init(".4,.3") 

and play around with starting points.

I hope to have clarified,

Gabriele

[Mark Pham]

Hi Gabriele,

Thanks a lot! This clarifies all of my issues! I expect -prodest- would be more and more popular soon.

Best regards,
Mark

[Klodi Gjini]

Dear Gabriele, I have one question: Is it possible that from the prodest estimation (met(lp)) to get negative values of tfp? can the measure of total factor productivity (the residual is in log) be negative?

[Gabriele Rovigatti]

Dear Klodi,

given that predict in prodest yields - by default - omega = log(TFP), let me differentiate between three cases:

1) TFP < 0. This, by construction, would yield omega = log(TFP) --> undefined logarithm (that is, you would have a missing value as a result in Stata).
2) 0 < TFP < 1 --> omega = log(TFP) < 0. In this case, you would obtain a negative value of omega.
3) TFP > 1 --> omega = log(TFP) > 0. In this case, you'd have a positive value of the variable.

All the above are possible cases, although 1) and 2) are extreme events. Let me stress that they are very unusual, and in case you figure that it is your case, you should check carefully the data, the unit of measure, and your model in general.

I hope to have clarified,

Gabriele

[Agnieszka Matulska]

Dear Gabrielle,

I used prodest procedure in order to estimate TFP using LP method with ACF correction.

prodest lnva, free(lnL) proxy(lninput) state(lnK) met(lp) reps(50) poly(3) acf valueadded fsresiduals(aa)

In order to estimate TFP I applied the following command: predict lnTFP, resid and then generate TFP_LP=exp(lnTFP_LP)

In the file with the results I got data on ln TFP, TFP and aa as residuls. Howeevr, when I calcultae the equatation lnva= alfa*lnL + beta*lnK+lnTFP+aa it turned out that there is no equaliity betwee lef and right side of equation. The equaliity exists when I do not take the residulas.

As a result I am not sure what I got for model. I would be very grateful if you could explain that to e.

Agnieszka

[Agnieszka Matulska]

What commands should be used in order to estimate TFP/lnTFP and residulas?

[Gabriele Rovigatti]

Dear Agnieszka,

in order to estimate the TFP after prodest you are running the correct command. Indeed, by construction lnva= alfa*lnL + beta*lnK+lnTFP , and the first-stage residuals (aa) should be left out of the equation.

I hope to have clarified,

Gabriele

[Agnieszka Matulska]

Thank you very much for help. But what about the command:

predict xx, omega

What kind of data do I receive from that command?

Agnieszka

[Gabriele Rovigatti]

Dear Agnieszka,

I am sorry for the previous message: I have probably been slack in answering and did not read correctly your issue.

In prodest the TFP is phi_it - hat{y}, so you should run predict lnTFP, omega to yield the result you need. By construction, predict newvar, resid yields y - hat{y}. Please refer to the helpfile (type help prodest_p) for a full explanation of available post-estimation commands.

Best,
Gabriele

[Agnieszka Matulska]

Dear Gabrielle,

I am really happy that you confirmed that I used proper command.

Thank you very much for help!!!
Agnieszka

[Michael Zimmermann]

Dear Gabrielle,

I am using the -prodest- command for a project and like all the others I find it extremely helpful and easy to use, thanks for providing the code!
Using paneldata

(two periods with gap)

I was able to implement the op- and lp-method in an instance.
My questions regard the ACF-correction implemented in -prodest-.

The code is basically:

prodest lny, free(lnl lnm lninv) state(lnc) proxy(lne) met(lp) va poly(4) reps(250) id(panelid) t(year)

When I now add the "acf"-option I receive the error message "missing values returned by evaluator" when using "nm"-optimizer and "Initial values not feasible" when using all the other optimizers.

1) Is -prodest- able to implement the ACF-correction if the panel consists of just two periods? (since OP and LP worked perfectly)

2) If so, what should I try to overcome the issue with the "missing values returned by evaluator"?


It would be a pleasure to hear from you.

Sincerely,

Michael

[Gabriele Rovigatti]

Dear Michael,

the "missing value returned by evaluator" error might be due to several reasons, hence I am not able to give you a definitive answer without looking at the data / code.

I would ask you to send me a reproducible example of the error (i.e., a sample of the data and a code snippet that reproduces the "strange" behavior on the sample data) to [email protected] .

Best,

Gabriele

[Yihao Chen]

Dear Gabrielle,
Thanks a lot for your great effort on the prodest package. It do help me a lot.
I have a problem now. When using the command
prodest ln_va, free(ln_wks ln_mat) proxy(ln_inv) state(ln_cap) control(ownrep age) va reps(50) method(op) id(stkcd) t(year)
predict lntfpop,resid
to calculate the tfp with the same dataset for many times, I found that the predictions were quite unstable. Maybe 2.x for this time and -150.x the next time.
I am lack of experience in using Stata and emperical research. So really hope a possible explanation and solution of this problem from you.
Sincerely,
Yihao