Code to Use Estimated Coefficient

[Issa Almaharmeh]

Hi All,
I have the following model

TA/LTA = α1 (1/LTA) + α2 (ΔREV/LTA) + α3 (PPE/LTA) + ε

which I need to run it for each firm in the same industry_id in each year. After running the this model I need to use the estimated coefficient of this model a1,a2, and ,a3 to calculate other variable as follow

NDA = α1 (1/LTA) + α2 (ΔREV/LTA) + α3 (PPE/LTA)

, so the new variable is NDA.
sorry if I wasn't clear but I'm tried my best.
Thank you in advance for your comments and suggestion

[Clyde Schechter]

If all you need to do is calculate NDA, there is no need to save the estimated coefficients separately, because -predict- will calculate NDA for you. If you do need the coefficients themselves for other reasons, see the help and manual sections for the -statsby- command. But to just get NDA you can do this (I'm calling your variables y x1 x2 and x3 for short, and I assume industry_id is a numeric variable, not a string):

levelsof industry_id, local(ids)
levelsof year, local(years)
gen nda = .
foreach i of local ids {
     foreach y of local years {
          regress y x1 x2 x3 if year == `y' & industry_id == `i'
          predict xb, xb
          replace nda = xb if year == `y' & industry_id == `i'
          drop xb
     }
}

[Issa Almaharmeh]

Dear Clyde
I run your code but there is as error message that no observation

no observations
r(2000);

.

[Nick Cox]

This could be a problem with missing data or with variables that should be numeric but are string. You may need to tell us more about your dataset to get more detailed advice.

Also, you did not reply to the earlier comment that you need enough observations for the regression to run, i.e. replicates for each firm-year combination.

http://www.statalist.org/forums/foru...-s-coefficient

Preceding

regress y x1 x2 x3 if year == `y' & industry_id == `i'

with

noi count if !missing(y, x1, x2, x3) & if year == `y' & industry_id == `i'

will show you how many observations there are for each regression you are trying.

[Issa Almaharmeh]

Dear Nick
I have 6376 firm year observations from 59 industry. In addition my firm_id and industry_id variables are numeric.
Hope this will help.

[Nick Cox]

That rules out a problem with string variables. Otherwise, that says nothing whatsoever about individual combinations of firm and year.

[Issa Almaharmeh]

This is a sample of my data. it is unbalanced panel of 6376 observations for about 650 firms frrom 59 industry for the period from 1990-2013.

y	x1	x2	x3	year	industry_id	firm_id
0.016866	5.44E-06	-0.04788	0.570729	2007	33	2002
0.167458	4.24E-05	-0.18187	0.060609	2010	44	2003
-0.00046	4.57E-05	0.155479	0.071096	2011	44	2003
-0.1809	3.63E-05	-0.09731	0.064317	2012	44	2003
-0.23769	5.24E-05	-0.13225	0.095747	2013	44	2003
				1997	57	2005
-0.1368	4.09E-06	0.205387	0.003296	2006	13	2011
0.255565	0.000824	0.79967	0.299258	1999	73	2012
-0.30504	0.000191	0.479206	0.137925	2000	73	2012
-0.41297	2.93E-05	0.112503	0.024783	2001	73	2012
-0.02425	6.73E-05	0	0.029903	2009	10	2014
-0.1044	0.000053	0	0.026444	2011	10	2014
-0.61345	1.32E-05	0.00286	0.917697	2006	49	2016
0.796652	0.000138	0.553465	0.113432	2002	87	2018
0.74341	7.96E-05	0.538345	0.128375	2003	87	2018
0.366284	1.26E-05	0.251456	0.039	2006	87	2018
0.503076	5.99E-06	0.384956	0.062444	2007	87	2018
0.544914	3.79E-06	0.345077	0.016366	2008	87	2018
0.05812	5.13E-05	0.174823	0.107623	2007	73	2024
0.074219	4.13E-05	0.219148	0.10247	2008	73	2024
0.144096	3.41E-05	0.479478	0.101892	2009	73	2024
0.011866	1.58E-05	0.073173	0.055827	2010	73	2024
-0.10689	0.000014	-0.01466	0.061827	2011	73	2024
-0.27348	4.19E-06	-0.00067	1.980653	2010	10	2027

[Nick Cox]

I see here (e.g.) one observation in which year is 2008 and industry id is 73, one with 2009 and 73, and so on. If that's correct. You can't run your regression for such combinations and you need more defensive code.

[Issa Almaharmeh]

could you help me on writing the codeplease?

[Nick Cox]

I'd personally not to want fit a multiple regression with 3 predictors to fewer than say 30 observations. That being so, a code sketch based on Clyde's code is

levelsof industry_id, local(ids)
levelsof year, local(years)
gen nda = .
gen nobs = .

foreach i of local ids {
     foreach y of local years {
          count if !missing(y, x1, x2, x3) &  year == `y' & industry_id == `i'
          replace nobs = r(N) if year == `y' & industry_id == `i'
    
          if r(N) >= 30 {
              regress y x1 x2 x3 if year == `y' & industry_id == `i'
              predict xb, xb
              replace nda = xb if year == `y' & industry_id == `i'
              drop xb
         }
     }
}

[Issa Almaharmeh]

Can I add more clarification here. I want to run the regression for each industry in every year (so the regression will estimate industry level coefficients) then we need to use these coefficients to calculate NDA for each firm in each year(so the firms with the same industry_id will have the same coefficients in the same year) .*** also there is a condition here to run the yearly regression for each industry that must be at least 6 firms in each industry in each year.
sorry again if the thing not clear but I did my best.

[Clyde Schechter]

Can I add more clarification here. I want to run the regression for each industry in every year (so the regression will estimate industry level coefficients) then we need to use these coefficients to calculate NDA for each firm in each year(so the firms with the same industry_id will have the same coefficients in the same year)

The code that Nick gives in #10 will do this.

.*** also there is a condition here to run the yearly regression for each industry that must be at least 6 firms in each industry in each year.

So you need to count the number of firms in each industry in each year and then then restrict the code to those industries where the number is at least 6. I assume that each combination of industry_id, firm_id, and year occurs only once in your data, which is true in your sample data. If it is not true in your data as a whole, a slightly different approach would be needed.

// VERIFY EACH COMBINATION OF industry_id, firm_id, & year
// OCCURS ONLY ONCE TO AVOID DOUBLE COUNTING
isid industry_id year firm_id, sort 

// COUNT NUMBER OF FIRMS IN EACH INDUSTRY IN EACH YEAR
by industry_id year( firm_id): gen firm_count _this_year= _N

// IDENTIFY SMALLEST NUMBER OF FIRMS IN ANY YEAR FOR EACH INDUSTRY
by industry_id (firm_count_this_year), sort: gen least_number_firms_any_year = firm_count_this_year[1]

Now, modify the code from #10 to add this restriction to the guard condition for the regression:

levelsof industry_id, local(ids)
levelsof year, local(years)
gen nda = .
gen nobs = .

foreach i of local ids {
     foreach y of local years {
          count if !missing(y, x1, x2, x3) &  year == `y' & industry_id == `i'
          replace nobs = r(N) if year == `y' & industry_id == `i'
    
          if r(N) >= 30 & least_number_firms_any_year >= 6 {
              regress y x1 x2 x3 if year == `y' & industry_id == `i'
              predict xb, xb
              replace nda = xb if year == `y' & industry_id == `i'
              drop xb
         }
     }
}

[Issa Almaharmeh]

I received this error message

factor variables and time-series operators not allowed
r(101);

when I tried to calculate firm count in each industry for each year by running the following code,
by industry_id year( firm_id): gen firm_count _this_year= _N

[Clyde Schechter]

Oh, sorry, there was a typo. There needs to be a space between year and (firm_id). (And no space needed between the open parenthesis and firm_id).

[Issa Almaharmeh]

After the correction I received the following error

too many variables specified
r(103);

could you help me on this please.