[Jose-Luis Herrera]

Dear Gabriele,

I am new on PTF estimation. I am using the ACF aproach and began with the acfest comand which use GMM. By construction the command use the lags of the regresors for the control function of the second stage and so on some observations are lost. Instead, with prodest ACF is an option included on LP or OP methods, and all the paneldata set is consider for the estimation. Why occur that difference? is the ACF correction in prodest different than ACF aproach? Thank you very much in advance.

prodest prod, free(labor) proxy(m) state(capital) poly(3) met(lp) valueadded reps(50) acf



acfest prod, free(labor) state(capital) proxy(m) i(iruc) t(year) va


Best, Jose-Luis.

[Gabriele Rovigatti]

Dear Jose-Luis,

the analysis sample in the two cases is the same. The reported number is different, though, because prodest reports the complete dataset, including the lagged values, whereas acfest discount that - to be completely honest, I think that acfest approach is more correct, I will try to release a new version of prodest that fixes this.

Thank you very much,

Gabriele

[Enjie Ma]

Dear Gabriele,

Again thank you for such a helpful package!

I have some questions on manually calculating TFP (omega) under ACF translog with control (i have to do it manually, since the predict function does not seem to work under translog with controls). More specifically, if one should subtract controls (beta_controls * controls) when calculating omega.

This is my main estimation code:

prodest log_y , free(log_lab) state(log_k) proxy(log_material) va att trans poly(3) met(lp) acf id(id) t(closing_year) control(di* dy*) fsresiduals(fr_res_acf_translog_`i')
replace fr_res_acf_translog = fr_res_acf_translog_`i' if industry == `i'
replace output_elas_labor_translog = _b[log_lab] + 2*log_lab*_b[var_1_1] + _b[var_1_2]*log_k if industry == `i'

... code on predicting markup...

replace tfp_translog = log_y - ( _b[log_lab]*log_lab+ _b[log_k]*log_k+ _b[var_1_1]*log_lab*log_lab+_b[var_2_2]*log_k*log_k+_b[var_1_2]*log_lab*log_k+predicted_fe)- fr_res_acf_translog 

predicted_fe is \beta*(di)* di + \beta*(dy) * dy. I am extracting each coefficient of beta_control's individually , multiplying it by the controls, and summing it across all controls.

My questions:

• Definition of TFP (I assume log_y - fr_res_acf_translog is the phi, and the beta_l*l + beta_k*k ... is the hat{y}. TFP = omega = phi - hat(y) = (log_y - fr_res_acf_translog) -(beta_l*log_lab + beta_k*log_k) ) Is this definition correct? If so, my impression is that I don't need to include the controls in the TFP calculation, as the controls' information is included in the first stage error already? I am a bit confused here.

• Is there any convenient way to extract the "predict_fe"'s ? The beta_control * control's.
• for translog coefficient, i assume var_1_1 corresponds to beta_l_l and var_2_2 correspond to beta_k_k? What does 1 and 2 represent in general? Free and state?
• Under translog, for firms with extremely small labor costs (hence negative log value), the labor elasticity, and therefore markup, can become negative. How should I think about negative elasticity? Should I remove them post or pre estimations?

I am showing my estimation result below:

Thank you so much!

Best regards,
Enjie