*This dofile estimates AD Probits on 5-6 digit NAICS data for AER revision.
 #delimit ;
clear matrix;
clear;
set mem 700m;
set more off;


* set path;
* (this is the only time that the absolute path is set in this code);

cd "C:\mypath" ;

log using Estimation-Table-2-05-01-2012, replace;

/* ***************************************************************************** */
/* ** TABLE 2: Specification (1): Baseline model  ** */
/* ***************************************************************************** */

clear;
use "SETA_estim_final_05-01-2012.dta";

/* RESCALE VARIABLES */

replace gr_imp_L1=gr_imp_L1/10000;
replace sd_imp_gr= sd_imp_gr/100;
replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum/1000;
replace chg_rxr_L1= chg_rxr_L1/1000;

*Generate the interacted continuous variables;
gen MDXS_imp= ln_inv_es_id_elast_sum*gr_imp_L1;


*As a dprobit; 
dprobit AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, robust;			

mfx compute;		

*As a probit, with interaction; 

probit AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, robust;								

keep if e(sample)==1; 

sum AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1; 											

*Calculate marginal effects;
*To verify code, start by recreating dprobit coefficients;
predict x_beta1, xb;
egen av_xbeta1=mean(x_beta1); 
scalar phi1=normalden(av_xbeta1); 
scalar list phi1; 



matrix marg=phi1*e(b); 
matrix list marg; 



*Next, calculate the interaction terms we’re interested in; 

egen av_imp_gr=mean(gr_imp_L1);													 									
egen av_elast=mean(ln_inv_es_id_elast_sum);													 							

gen me_imports= (_b[gr_imp_L1]+_b[MDXS_imp]*av_elast)*phi1;
gen me_elast= (_b[ln_inv_es_id_elast_sum]+_b[MDXS_imp]*av_imp_gr)*phi1;

/* ***************************************************************************** */
/* THESE ARE THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum me_imports;
sum  me_elast; 

*Calculate std errors of marginal effects; 
*To verify code, start by recreating dprobit coefficients;
*Note that when working with the probit command, “robust,” the vce matrix is robust and the associated std errors are the “robust” std errors from dprobit; 

*define the VCE matrix;
matrix varcov=e(V);

matrix list varcov; 


*extract the diagonal of the vce matrix;
matrix vc_row=vecdiag(varcov);
*construct a new matrix with zeros on the off-diagonal;
matrix vc_diag=diag(vc_row);
*take the square root of the diagonal elements; 
matrix std_err=cholesky(vc_diag);
*scale up to marginal effects; 
matrix std_err_marg=phi1*std_err; 
matrix std_err_scaled=vecdiag(std_err_marg);
matrix list std_err_scaled; 

matrix list varcov; 

gen var_me_imp =varcov[1,1]+(varcov[4,4]*(av_elast^2)) +2*av_elast*varcov[1,4];	

gen sqrt_var_imp=var_me_imp^(.5); 
gen std_err_me_imp=sqrt_var_imp*phi1; 

gen var_me_elast=varcov[3,3] +(varcov[4,4]*(av_imp_gr^2))+2*av_imp_gr*varcov[3,4];
gen sqrt_var_elast=var_me_elast^(.5);
gen std_err_me_elast=sqrt_var_elast*phi1;

/* ***************************************************************************** */
/* THESE ARE THE STANDARD ERRORS OF THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum std_err_me_imp; 
sum std_err_me_elast; 




************************************************************************************************;
*PREDICTION: quantify the effect on the prob of a one std dev shock for interaction terms;
************************************************************************************************;
 
*Calculate the predicted value at the mean of the dataset;
predict x_beta_1;
sum x_beta_1;
 
*Generate local variables equal to the mean and std dev of gr_imp_L1;
egen gr_imp_L1_mean= mean(gr_imp_L1);
egen gr_imp_L1_sd=sd(gr_imp_L1);
gen gr_imp_L1_saved=gr_imp_L1;
replace gr_imp_L1= gr_imp_L1_mean + gr_imp_L1_sd;

*Generate local variables equal to the mean and std dev of the elasticities;
egen ln_inv_es_id_elast_sum_mean= mean(ln_inv_es_id_elast_sum);
egen ln_inv_es_id_elast_sum_sd=sd(ln_inv_es_id_elast_sum);
 
*Part 1: Consider the effect on the probability of a one std deviation shock to import growth; 
predict x_beta_2;
sum x_beta_2;

gen  MDXS_imp_saved=MDXS_imp;
replace MDXS_imp= ln_inv_es_id_elast_sum*(gr_imp_L1_mean+gr_imp_L1_sd);

/* ***************************************************************************** */
/* THIS (x_beta_3) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO IMP GROWTH */
 
predict x_beta_3;
sum x_beta_3;


*Part 2: Consider the effect of a one standard deviation shock to the elasticity measure;
 
*First, change those values back to the means;
replace gr_imp_L1=gr_imp_L1_saved;
replace MDXS_imp=MDXS_imp_saved;

/* ***************************************************************************** */
*Check that this (x_beta_4) is the same as x_beta_1 so the original data is reloaded into memory;
predict x_beta_4;
sum x_beta_4;

replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum_mean + ln_inv_es_id_elast_sum_sd;

predict x_beta_5;
sum x_beta_5;
 
replace MDXS_imp= gr_imp_L1*(ln_inv_es_id_elast_sum_mean+ln_inv_es_id_elast_sum_sd);

/* ***************************************************************************** */
/* THIS  (x_beta_6) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO THE ELASTICITY MEASURE */
 
predict x_beta_6;
sum x_beta_6;


 



/* ***************************************************************************** */
/* ** TABLE 2: Specification (2): Check on Baseline model, alternate definition of elasticities as = 1/(xs+id) ** */
/* ***************************************************************************** */

#delimit;
clear;
use "SETA_estim_final_05-01-2012.dta";

/* RESCALE VARIABLES */

replace gr_imp_L1=gr_imp_L1/10000;
replace sd_imp_gr= sd_imp_gr/100;
replace inv_es_id_elast_sum= inv_es_id_elast_sum/100;
replace chg_rxr_L1= chg_rxr_L1/1000;


*Generate the interacted continuous variables;
gen MDXS_imp= inv_es_id_elast_sum*gr_imp_L1;


*As a dprobit; 
dprobit AD_imposed  gr_imp_L1 sd_imp_gr inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, robust;								

mfx compute;		

*As a probit, with interaction; 

probit AD_imposed  gr_imp_L1 sd_imp_gr inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, robust;								

keep if e(sample)==1; 

sum AD_imposed  gr_imp_L1 sd_imp_gr inv_es_id_elast_sum MDXS_imp  chg_rxr_L1; 											

*Calculate marginal effects;
*To verify code, start by recreating dprobit coefficients;
predict x_beta1, xb;
egen av_xbeta1=mean(x_beta1); 
scalar phi1=normalden(av_xbeta1); 
scalar list phi1; 



matrix marg=phi1*e(b); 
matrix list marg; 



*Next, calculate the interaction terms we’re interested in; 

egen av_imp_gr=mean(gr_imp_L1);													 									
egen av_elast=mean(inv_es_id_elast_sum);													 							

gen me_imports= (_b[gr_imp_L1]+_b[MDXS_imp]*av_elast)*phi1;
gen me_elast= (_b[inv_es_id_elast_sum]+_b[MDXS_imp]*av_imp_gr)*phi1;

/* ***************************************************************************** */
/* THESE ARE THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum me_imports;
sum  me_elast; 

*Calculate std errors of marginal effects; 
*To verify code, start by recreating dprobit coefficients;
*Note that when working with the probit command, “robust,” the vce matrix is robust and the associated std errors are the “robust” std errors from dprobit; 

*define the VCE matrix;
matrix varcov=e(V);

matrix list varcov; 


*extract the diagonal of the vce matrix;
matrix vc_row=vecdiag(varcov);
*construct a new matrix with zeros on the off-diagonal;
matrix vc_diag=diag(vc_row);
*take the square root of the diagonal elements; 
matrix std_err=cholesky(vc_diag);
*scale up to marginal effects; 
matrix std_err_marg=phi1*std_err; 
matrix std_err_scaled=vecdiag(std_err_marg);
matrix list std_err_scaled; 

matrix list varcov; 

gen var_me_imp =varcov[1,1]+(varcov[4,4]*(av_elast^2)) +2*av_elast*varcov[1,4];	

gen sqrt_var_imp=var_me_imp^(.5); 
gen std_err_me_imp=sqrt_var_imp*phi1; 

gen var_me_elast=varcov[3,3] +(varcov[4,4]*(av_imp_gr^2))+2*av_imp_gr*varcov[3,4];
gen sqrt_var_elast=var_me_elast^(.5);
gen std_err_me_elast=sqrt_var_elast*phi1;

/* ***************************************************************************** */
/* THESE ARE THE STANDARD ERRORS OF THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum std_err_me_imp; 
sum std_err_me_elast; 

************************************************************************************************;
*PREDICTION: quantify the effect on the prob of a one std dev shock for interaction terms;
************************************************************************************************;
 
*Calculate the predicted value at the mean of the dataset;
predict x_beta_1;
sum x_beta_1;
 
*Generate local variables equal to the mean and std dev of gr_imp_L1;
egen gr_imp_L1_mean= mean(gr_imp_L1);
egen gr_imp_L1_sd=sd(gr_imp_L1);
gen gr_imp_L1_saved=gr_imp_L1;
replace gr_imp_L1= gr_imp_L1_mean + gr_imp_L1_sd;

*Generate local variables equal to the mean and std dev of the elasticities;
egen inv_es_id_elast_sum_mean= mean(inv_es_id_elast_sum);
egen inv_es_id_elast_sum_sd=sd(inv_es_id_elast_sum);
 
*Part 1: Consider the effect on the probability of a one std deviation shock to import growth; 
predict x_beta_2;
sum x_beta_2;

gen  MDXS_imp_saved=MDXS_imp;
replace MDXS_imp= inv_es_id_elast_sum*(gr_imp_L1_mean+gr_imp_L1_sd);

/* ***************************************************************************** */
/* THIS (x_beta_3) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO IMP GROWTH */
 
predict x_beta_3;
sum x_beta_3;


*Part 2: Consider the effect of a one standard deviation shock to the elasticity measure;
 
*First, change those values back to the means;
replace gr_imp_L1=gr_imp_L1_saved;
replace MDXS_imp=MDXS_imp_saved;

/* ***************************************************************************** */
*Check that this (x_beta_4) is the same as x_beta_1 so the original data is reloaded into memory;
predict x_beta_4;
sum x_beta_4;

replace inv_es_id_elast_sum= inv_es_id_elast_sum_mean + inv_es_id_elast_sum_sd;

predict x_beta_5;
sum x_beta_5;
 
replace MDXS_imp= gr_imp_L1*(inv_es_id_elast_sum_mean+inv_es_id_elast_sum_sd);

/* ***************************************************************************** */
/* THIS  (x_beta_6) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO THE ELASTICITY MEASURE */
 
predict x_beta_6;
sum x_beta_6;



/* ***************************************************************************** */
/* ** TABLE 2: Specification (3): Baseline model, subsample after removing "outlier" XS and ID elasticities above and below 5% threshold ** */
/* ***************************************************************************** */
#delimit;
clear;
use "SETA_estim_final_05-01-2012.dta";


/* RESCALE VARIABLES */

replace gr_imp_L1=gr_imp_L1/10000;
replace sd_imp_gr= sd_imp_gr/100;
replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum/1000;
replace chg_rxr_L1= chg_rxr_L1/1000;

*Generate the interacted continuous variables;

gen MDXS_imp= ln_inv_es_id_elast_sum*gr_imp_L1;


dprobit AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, robust;
keep if e(sample)==1;

sum  inv_id_elast_med, detail; 
drop if inv_id_elast_med>.810888 | inv_id_elast_med<.0922983; 

sum  inv_es_elast_med, detail; 
drop if inv_es_elast_med>5.518096 | inv_es_elast_med<.0067208; 



*As a dprobit; 
dprobit AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, robust;								

mfx compute;		

*As a probit, with interaction; 

probit AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, robust;								

keep if e(sample)==1; 

sum AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1; 											

*Calculate marginal effects;
*To verify code, start by recreating dprobit coefficients;
predict x_beta1, xb;
egen av_xbeta1=mean(x_beta1); 
scalar phi1=normalden(av_xbeta1); 
scalar list phi1; 



matrix marg=phi1*e(b); 
matrix list marg; 



*Next, calculate the interaction terms we’re interested in; 

egen av_imp_gr=mean(gr_imp_L1);													 									
egen av_elast=mean(ln_inv_es_id_elast_sum);													 							

gen me_imports= (_b[gr_imp_L1]+_b[MDXS_imp]*av_elast)*phi1;
gen me_elast= (_b[ln_inv_es_id_elast_sum]+_b[MDXS_imp]*av_imp_gr)*phi1;

/* ***************************************************************************** */
/* THESE ARE THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum me_imports;
sum  me_elast; 

*Calculate std errors of marginal effects; 
*To verify code, start by recreating dprobit coefficients;
*Note that when working with the probit command, “robust,” the vce matrix is robust and the associated std errors are the “robust” std errors from dprobit; 

*define the VCE matrix;
matrix varcov=e(V);

matrix list varcov; 


*extract the diagonal of the vce matrix;
matrix vc_row=vecdiag(varcov);
*construct a new matrix with zeros on the off-diagonal;
matrix vc_diag=diag(vc_row);
*take the square root of the diagonal elements; 
matrix std_err=cholesky(vc_diag);
*scale up to marginal effects; 
matrix std_err_marg=phi1*std_err; 
matrix std_err_scaled=vecdiag(std_err_marg);
matrix list std_err_scaled; 

matrix list varcov; 

gen var_me_imp =varcov[1,1]+(varcov[4,4]*(av_elast^2)) +2*av_elast*varcov[1,4];	

gen sqrt_var_imp=var_me_imp^(.5); 
gen std_err_me_imp=sqrt_var_imp*phi1; 

gen var_me_elast=varcov[3,3] +(varcov[4,4]*(av_imp_gr^2))+2*av_imp_gr*varcov[3,4];
gen sqrt_var_elast=var_me_elast^(.5);
gen std_err_me_elast=sqrt_var_elast*phi1;

/* ***************************************************************************** */
/* THESE ARE THE STANDARD ERRORS OF THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum std_err_me_imp; 
sum std_err_me_elast; 

************************************************************************************************;
*PREDICTION: quantify the effect on the prob of a one std dev shock for interaction terms;
************************************************************************************************;
 
*Calculate the predicted value at the mean of the dataset;
predict x_beta_1;
sum x_beta_1;
 
*Generate local variables equal to the mean and std dev of gr_imp_L1;
egen gr_imp_L1_mean= mean(gr_imp_L1);
egen gr_imp_L1_sd=sd(gr_imp_L1);
gen gr_imp_L1_saved=gr_imp_L1;
replace gr_imp_L1= gr_imp_L1_mean + gr_imp_L1_sd;

*Generate local variables equal to the mean and std dev of the elasticities;
egen ln_inv_es_id_elast_sum_mean= mean(ln_inv_es_id_elast_sum);
egen ln_inv_es_id_elast_sum_sd=sd(ln_inv_es_id_elast_sum);
 
*Part 1: Consider the effect on the probability of a one std deviation shock to import growth; 
predict x_beta_2;
sum x_beta_2;

gen  MDXS_imp_saved=MDXS_imp;
replace MDXS_imp= ln_inv_es_id_elast_sum*(gr_imp_L1_mean+gr_imp_L1_sd);

/* ***************************************************************************** */
/* THIS (x_beta_3) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO IMP GROWTH */
 
predict x_beta_3;
sum x_beta_3;


*Part 2: Consider the effect of a one standard deviation shock to the elasticity measure;
 
*First, change those values back to the means;
replace gr_imp_L1=gr_imp_L1_saved;
replace MDXS_imp=MDXS_imp_saved;

/* ***************************************************************************** */
*Check that this (x_beta_4) is the same as x_beta_1 so the original data is reloaded into memory;
predict x_beta_4;
sum x_beta_4;

replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum_mean + ln_inv_es_id_elast_sum_sd;

predict x_beta_5;
sum x_beta_5;
 
replace MDXS_imp= gr_imp_L1*(ln_inv_es_id_elast_sum_mean+ln_inv_es_id_elast_sum_sd);

/* ***************************************************************************** */
/* THIS  (x_beta_6) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO THE ELASTICITY MEASURE */
 
predict x_beta_6;
sum x_beta_6;



/* ***************************************************************************** */
/* ** TABLE 2: Specification (4): Baseline model, Subsample on only the top 10 US import sources ** */
/* ***************************************************************************** */

clear;
use "SETA_estim_final_05-01-2012.dta";


#delimit;
keep if country=="Canada" | country=="Japan" | country=="Mexico" |  country=="China"   | country=="Germany" | country=="Taiwan" | country=="United Kingdom"
		| country=="South Korea" | country=="France" | country=="Italy" | country=="Malaysia" ;  


/* RESCALE VARIABLES */

replace gr_imp_L1=gr_imp_L1/10000;
replace sd_imp_gr= sd_imp_gr/100;
replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum/1000;
replace chg_rxr_L1= chg_rxr_L1/1000;

*Generate the interacted continuous variables;

gen MDXS_imp= ln_inv_es_id_elast_sum*gr_imp_L1;


*As a dprobit; 
dprobit AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, robust;								

mfx compute;		

*As a probit, with interaction; 

probit AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, robust;								

keep if e(sample)==1; 

sum AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1; 											

*Calculate marginal effects;
*To verify code, start by recreating dprobit coefficients;
predict x_beta1, xb;
egen av_xbeta1=mean(x_beta1); 
scalar phi1=normalden(av_xbeta1); 
scalar list phi1; 



matrix marg=phi1*e(b); 
matrix list marg; 



*Next, calculate the interaction terms we’re interested in; 

egen av_imp_gr=mean(gr_imp_L1);													 									
egen av_elast=mean(ln_inv_es_id_elast_sum);													 							

gen me_imports= (_b[gr_imp_L1]+_b[MDXS_imp]*av_elast)*phi1;
gen me_elast= (_b[ln_inv_es_id_elast_sum]+_b[MDXS_imp]*av_imp_gr)*phi1;

/* ***************************************************************************** */
/* THESE ARE THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum me_imports;
sum  me_elast; 

*Calculate std errors of marginal effects; 
*To verify code, start by recreating dprobit coefficients;
*Note that when working with the probit command, “robust,” the vce matrix is robust and the associated std errors are the “robust” std errors from dprobit; 

*define the VCE matrix;
matrix varcov=e(V);

matrix list varcov; 


*extract the diagonal of the vce matrix;
matrix vc_row=vecdiag(varcov);
*construct a new matrix with zeros on the off-diagonal;
matrix vc_diag=diag(vc_row);
*take the square root of the diagonal elements; 
matrix std_err=cholesky(vc_diag);
*scale up to marginal effects; 
matrix std_err_marg=phi1*std_err; 
matrix std_err_scaled=vecdiag(std_err_marg);
matrix list std_err_scaled; 

matrix list varcov; 

gen var_me_imp =varcov[1,1]+(varcov[4,4]*(av_elast^2)) +2*av_elast*varcov[1,4];	

gen sqrt_var_imp=var_me_imp^(.5); 
gen std_err_me_imp=sqrt_var_imp*phi1; 

gen var_me_elast=varcov[3,3] +(varcov[4,4]*(av_imp_gr^2))+2*av_imp_gr*varcov[3,4];
gen sqrt_var_elast=var_me_elast^(.5);
gen std_err_me_elast=sqrt_var_elast*phi1;

/* ***************************************************************************** */
/* THESE ARE THE STANDARD ERRORS OF THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum std_err_me_imp; 
sum std_err_me_elast; 

************************************************************************************************;
*PREDICTION: quantify the effect on the prob of a one std dev shock for interaction terms;
************************************************************************************************;
 
*Calculate the predicted value at the mean of the dataset;
predict x_beta_1;
sum x_beta_1;
 
*Generate local variables equal to the mean and std dev of gr_imp_L1;
egen gr_imp_L1_mean= mean(gr_imp_L1);
egen gr_imp_L1_sd=sd(gr_imp_L1);
gen gr_imp_L1_saved=gr_imp_L1;
replace gr_imp_L1= gr_imp_L1_mean + gr_imp_L1_sd;

*Generate local variables equal to the mean and std dev of the elasticities;
egen ln_inv_es_id_elast_sum_mean= mean(ln_inv_es_id_elast_sum);
egen ln_inv_es_id_elast_sum_sd=sd(ln_inv_es_id_elast_sum);
 
*Part 1: Consider the effect on the probability of a one std deviation shock to import growth; 
predict x_beta_2;
sum x_beta_2;

gen  MDXS_imp_saved=MDXS_imp;
replace MDXS_imp= ln_inv_es_id_elast_sum*(gr_imp_L1_mean+gr_imp_L1_sd);

/* ***************************************************************************** */
/* THIS (x_beta_3) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO IMP GROWTH */
 
predict x_beta_3;
sum x_beta_3;


*Part 2: Consider the effect of a one standard deviation shock to the elasticity measure;
 
*First, change those values back to the means;
replace gr_imp_L1=gr_imp_L1_saved;
replace MDXS_imp=MDXS_imp_saved;

/* ***************************************************************************** */
*Check that this (x_beta_4) is the same as x_beta_1 so the original data is reloaded into memory;
predict x_beta_4;
sum x_beta_4;

replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum_mean + ln_inv_es_id_elast_sum_sd;

predict x_beta_5;
sum x_beta_5;
 
replace MDXS_imp= gr_imp_L1*(ln_inv_es_id_elast_sum_mean+ln_inv_es_id_elast_sum_sd);

/* ***************************************************************************** */
/* THIS  (x_beta_6) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO THE ELASTICITY MEASURE */
 
predict x_beta_6;
sum x_beta_6;



/* ***************************************************************************** */
/* ** TABLE 2: Specification (5): Baseline model, Top 10 trading partners, logit model with CGM multiway clustering ** */
/* ***************************************************************************** */

clear;
use "SETA_estim_final_05-01-2012.dta";

#delimit;
keep if country=="Canada" | country=="Japan" | country=="Mexico" |  country=="China"   | country=="Germany" | country=="Taiwan" | country=="United Kingdom"
		| country=="South Korea" | country=="France" | country=="Italy" | country=="Malaysia" ;  
 

/* RESCALE VARIABLES */

replace gr_imp_L1=gr_imp_L1/10000;
replace sd_imp_gr= sd_imp_gr/100;
replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum/1000;
replace chg_rxr_L1= chg_rxr_L1/1000;


*Generate the interacted continuous variables;
gen MDXS_imp= ln_inv_es_id_elast_sum*gr_imp_L1;

*As a logit with interaction and multiway clustering; 

cgmlogit AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, cluster(country NAICS); 

*calculate marginal effects; 
mfx compute; 


keep if e(sample)==1; 

sum AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1; 

*Calculate marginal effects;
*To verify code, start by recreating mfx coefficients;
predict x_beta1, xb;
egen av_xbeta1=mean(x_beta1); 
scalar phi1= [ 1/(1+exp(-av_xbeta1)) ] *[1- (1/(1+exp(-av_xbeta1)))]; 
scalar list phi1; 

matrix marg=phi1*e(b); 
matrix list marg; 


*Next, calculate the interaction terms we’re interested in; 

egen av_imp_gr=mean(gr_imp_L1);
egen av_elast=mean(ln_inv_es_id_elast_sum);

gen me_imports= (_b[gr_imp_L1]+_b[MDXS_imp]*av_elast)*phi1;
gen me_elast= (_b[ln_inv_es_id_elast_sum]+_b[MDXS_imp]*av_imp_gr)*phi1;

/* ***************************************************************************** */
/* THESE ARE THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum me_imports;
sum  me_elast; 

*Calculate std errors of marginal effects; 

*define the VCE matrix;
matrix varcov=e(V);

matrix list varcov; 


*extract the diagonal of the vce matrix;
matrix vc_row=vecdiag(varcov);
*construct a new matrix with zeros on the off-diagonal;
matrix vc_diag=diag(vc_row);
*take the square root of the diagonal elements; 
matrix std_err=cholesky(vc_diag);
*scale up to marginal effects; 
matrix std_err_marg=phi1*std_err; 
matrix std_err_scaled=vecdiag(std_err_marg);
matrix list std_err_scaled; 

matrix list varcov; 

gen var_me_imp =varcov[1,1]+(varcov[4,4]*(av_elast^2)) +2*av_elast*varcov[1,4];

gen sqrt_var_imp=var_me_imp^(.5); 
gen std_err_me_imp=sqrt_var_imp*phi1; 

gen var_me_elast=varcov[3,3] +(varcov[4,4]*(av_imp_gr^2))+2*av_imp_gr*varcov[3,4];
gen sqrt_var_elast=var_me_elast^(.5);
gen std_err_me_elast=sqrt_var_elast*phi1;

/* ***************************************************************************** */
/* THESE ARE THE STANDARD ERRORS OF THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum std_err_me_imp; 
sum std_err_me_elast; 

************************************************************************************************;
*PREDICTION: quantify the effect on the prob of a one std dev shock for interaction terms;
************************************************************************************************;
 
*Calculate the predicted value at the mean of the dataset;
predict x_beta_1;
sum x_beta_1;
 
*Generate local variables equal to the mean and std dev of gr_imp_L1;
egen gr_imp_L1_mean= mean(gr_imp_L1);
egen gr_imp_L1_sd=sd(gr_imp_L1);
gen gr_imp_L1_saved=gr_imp_L1;
replace gr_imp_L1= gr_imp_L1_mean + gr_imp_L1_sd;

*Generate local variables equal to the mean and std dev of the elasticities;
egen ln_inv_es_id_elast_sum_mean= mean(ln_inv_es_id_elast_sum);
egen ln_inv_es_id_elast_sum_sd=sd(ln_inv_es_id_elast_sum);
 
*Part 1: Consider the effect on the probability of a one std deviation shock to import growth; 
predict x_beta_2;
sum x_beta_2;

gen  MDXS_imp_saved=MDXS_imp;
replace MDXS_imp= ln_inv_es_id_elast_sum*(gr_imp_L1_mean+gr_imp_L1_sd);

/* ***************************************************************************** */
/* THIS (x_beta_3) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO IMP GROWTH */
 
predict x_beta_3;
sum x_beta_3;


*Part 2: Consider the effect of a one standard deviation shock to the elasticity measure;
 
*First, change those values back to the means;
replace gr_imp_L1=gr_imp_L1_saved;
replace MDXS_imp=MDXS_imp_saved;

/* ***************************************************************************** */
*Check that this (x_beta_4) is the same as x_beta_1 so the original data is reloaded into memory;
predict x_beta_4;
sum x_beta_4;

replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum_mean + ln_inv_es_id_elast_sum_sd;

predict x_beta_5;
sum x_beta_5;
 
replace MDXS_imp= gr_imp_L1*(ln_inv_es_id_elast_sum_mean+ln_inv_es_id_elast_sum_sd);

/* ***************************************************************************** */
/* THIS  (x_beta_6) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO THE ELASTICITY MEASURE */
 
predict x_beta_6;
sum x_beta_6;



/* ***************************************************************************** */
/* ** TABLE 2: Alternative Specification (5): Baseline model, All trading partners, logit model with CGM multiway clustering ** */
/* ***************************************************************************** */

clear;
use "SETA_estim_final_05-01-2012.dta";


/* RESCALE VARIABLES */

replace gr_imp_L1=gr_imp_L1/10000;
replace sd_imp_gr= sd_imp_gr/100;
replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum/1000;
replace chg_rxr_L1= chg_rxr_L1/1000;


*Generate the interacted continuous variables;
gen MDXS_imp= ln_inv_es_id_elast_sum*gr_imp_L1;

*As a logit with interaction and multiway clustering; 

cgmlogit AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1, cluster(country NAICS); 

*calculate marginal effects; 
mfx compute; 


keep if e(sample)==1; 

sum AD_imposed  gr_imp_L1 sd_imp_gr ln_inv_es_id_elast_sum MDXS_imp  chg_rxr_L1; 

*Calculate marginal effects;
*To verify code, start by recreating mfx coefficients;
predict x_beta1, xb;
egen av_xbeta1=mean(x_beta1); 
scalar phi1= [ 1/(1+exp(-av_xbeta1)) ] *[1- (1/(1+exp(-av_xbeta1)))]; 
scalar list phi1; 

matrix marg=phi1*e(b); 
matrix list marg; 


*Next, calculate the interaction terms we’re interested in; 

egen av_imp_gr=mean(gr_imp_L1);
egen av_elast=mean(ln_inv_es_id_elast_sum);

gen me_imports= (_b[gr_imp_L1]+_b[MDXS_imp]*av_elast)*phi1;
gen me_elast= (_b[ln_inv_es_id_elast_sum]+_b[MDXS_imp]*av_imp_gr)*phi1;

/* ***************************************************************************** */
/* THESE ARE THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum me_imports;
sum  me_elast; 

*Calculate std errors of marginal effects; 

*define the VCE matrix;
matrix varcov=e(V);

matrix list varcov; 


*extract the diagonal of the vce matrix;
matrix vc_row=vecdiag(varcov);
*construct a new matrix with zeros on the off-diagonal;
matrix vc_diag=diag(vc_row);
*take the square root of the diagonal elements; 
matrix std_err=cholesky(vc_diag);
*scale up to marginal effects; 
matrix std_err_marg=phi1*std_err; 
matrix std_err_scaled=vecdiag(std_err_marg);
matrix list std_err_scaled; 

matrix list varcov; 

gen var_me_imp =varcov[1,1]+(varcov[4,4]*(av_elast^2)) +2*av_elast*varcov[1,4];

gen sqrt_var_imp=var_me_imp^(.5); 
gen std_err_me_imp=sqrt_var_imp*phi1; 

gen var_me_elast=varcov[3,3] +(varcov[4,4]*(av_imp_gr^2))+2*av_imp_gr*varcov[3,4];
gen sqrt_var_elast=var_me_elast^(.5);
gen std_err_me_elast=sqrt_var_elast*phi1;

/* ***************************************************************************** */
/* THESE ARE THE STANDARD ERRORS OF THE MARGINAL EFFECTS ESTIMATES INCLUDING THE DIRECT AND INDIRECT EFFECTS FOR THE VARIABLES OF INTEREST */

sum std_err_me_imp; 
sum std_err_me_elast; 

************************************************************************************************;
*PREDICTION: quantify the effect on the prob of a one std dev shock for interaction terms;
************************************************************************************************;
 
*Calculate the predicted value at the mean of the dataset;
predict x_beta_1;
sum x_beta_1;
 
*Generate local variables equal to the mean and std dev of gr_imp_L1;
egen gr_imp_L1_mean= mean(gr_imp_L1);
egen gr_imp_L1_sd=sd(gr_imp_L1);
gen gr_imp_L1_saved=gr_imp_L1;
replace gr_imp_L1= gr_imp_L1_mean + gr_imp_L1_sd;

*Generate local variables equal to the mean and std dev of the elasticities;
egen ln_inv_es_id_elast_sum_mean= mean(ln_inv_es_id_elast_sum);
egen ln_inv_es_id_elast_sum_sd=sd(ln_inv_es_id_elast_sum);
 
*Part 1: Consider the effect on the probability of a one std deviation shock to import growth; 
predict x_beta_2;
sum x_beta_2;

gen  MDXS_imp_saved=MDXS_imp;
replace MDXS_imp= ln_inv_es_id_elast_sum*(gr_imp_L1_mean+gr_imp_L1_sd);

/* ***************************************************************************** */
/* THIS (x_beta_3) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO IMP GROWTH */
 
predict x_beta_3;
sum x_beta_3;


*Part 2: Consider the effect of a one standard deviation shock to the elasticity measure;
 
*First, change those values back to the means;
replace gr_imp_L1=gr_imp_L1_saved;
replace MDXS_imp=MDXS_imp_saved;

/* ***************************************************************************** */
*Check that this (x_beta_4) is the same as x_beta_1 so the original data is reloaded into memory;
predict x_beta_4;
sum x_beta_4;

replace ln_inv_es_id_elast_sum= ln_inv_es_id_elast_sum_mean + ln_inv_es_id_elast_sum_sd;

predict x_beta_5;
sum x_beta_5;
 
replace MDXS_imp= gr_imp_L1*(ln_inv_es_id_elast_sum_mean+ln_inv_es_id_elast_sum_sd);

/* ***************************************************************************** */
/* THIS  (x_beta_6) IS THE FULL INTERRACTED EFFECT ON THE PROBABILITY OF A ONE STD DEV INCREASE TO THE ELASTICITY MEASURE */
 
predict x_beta_6;
sum x_beta_6;



log close;