How to this program

[Duong Le]

Hi all,

I am trying to write a program for my own analysis to save time because I have to repeat similar analyses numerous times. Since I am pretty new to Stata programming, I have tried a simple one as the followings:

cap program drop example
program  example
    args y x1 x2 x3
    di "`y' `x1' `x2' `x3'"
    
    qui reg `y' `x1' `x2' `x3', vce(robust)
        sum `y' if e(sample)
        sca m_y = r(mean)
        sca b = _b[`x1']
        sca se = _se[`x1']
        
        sca di m_y b se
end

    sysuse auto, clear
    example price headroom mpg rep78

That program works, but I now want to make it a bit more complicated. Suppose that I run two first stage, obtain first stage predicted values and then plug these predicted values into the second stage. Let y the outcome, z1 and z2 are instrumental variables, y1 and y2 are endogenous variables and x1 and x2 are exogenous variables. I want to write a program for the following regressions and store their associated estimates.

    reg y1 z1 x1, robust
        predict yhat1
        
    reg y2 z2 x2, robust
        predict yhat2
    
    reg y yhat1 yhat2 x1 x2, robust
        sum y if e(sample)
        sca m_y = r(mean)
        sca b = _b[yhat1]
        sca se = _se[yhat2]

Any help is much appreciated. Thank you.

[Duong Le]

I am sorry but any advice?

[Daniel Schaefer]

It's not really clear from your post what you are asking or what the problem is. Could you be more specific?

[Ken Chui]

Use rclass and it should be able to do what you need:

sysuse auto, clear

capture program drop example
program define example, rclass
    args y1 z1 x1 y2 z2 x2 y
    
    reg `y1' `z1' `x1', robust
    capture drop yhat1
    predict yhat1
    
    reg `y2' `z2' `x2', robust
    capture drop yhat2
    predict yhat2
    
    reg `y' yhat1 yhat2 `x1' `x2', robust
    return scalar b = _b[yhat1]
    return scalar se = _se[yhat2]
    
    sum `y' if e(sample)
    return scalar m_y = r(mean)
end

example price mpg rep78 headroom trunk length turn
return list

[Fei Wang]

Other than Ken's helpful code, I'd like to mention a theoretical issue. It seems that the OP plans to regress y on y1, y2, x1, and x2, where y1 and y2 are endogenous variables instrumented by z1 and z2. Then the first stage regressions should be

reg y1 z1 z2 x1 x2
reg y2 z1 z2 x1 x2

Another issue would be standard error adjustment. But I'll stop here, as Daniel has pointed out in #2, it's not quite clear what you were asking.

[Duong Le]

Other than Ken's helpful code, I'd like to mention a theoretical issue. It seems that the OP plans to regress y on y1, y2, x1, and x2, where y1 and y2 are endogenous variables instrumented by z1 and z2. Then the first stage regressions should be

reg y1 z1 z2 x1 x2
reg y2 z1 z2 x1 x2

Another issue would be standard error adjustment. But I'll stop here, as Daniel has pointed out in #2, it's not quite clear what you were asking.

Thank you Ken for the code. It works well.

Hi Fei, thank you for pointing out the theoretical issue. The code provided by Ken is what I wanted. I understand your code in #5, and a corresponding command in Stata could be: ivregress 2sls y (y1 y2 = z1 z2) x1 x2. In doing so, one would no need to adjust standard errors by "hands" because ivregress automatically adjusts standard errors for us. However, in my case, I need to run two separate first stages and then plug their predicted values into the second stage. Could you please guide me how to adjust standard error in my case?

Thank you.

[Fei Wang]

Thank you Ken for the code. It works well.

Hi Fei, thank you for pointing out the theoretical issue. The code provided by Ken is what I wanted. I understand your code in #5, and a corresponding command in Stata could be: ivregress 2sls y (y1 y2 = z1 z2) x1 x2. In doing so, one would no need to adjust standard errors by "hands" because ivregress automatically adjusts standard errors for us. However, in my case, I need to run two separate first stages and then plug their predicted values into the second stage. Could you please guide me how to adjust standard error in my case?

Thank you.

For your reference:
https://www.statalist.org/forums/for...tandard-errors
https://www.statalist.org/forums/for...-bootstrapping

[Duong Le]

For your reference:
https://www.statalist.org/forums/for...tandard-errors
https://www.statalist.org/forums/for...-bootstrapping

Thank Fei for the helpful references. I appreciate your help.

Following instructions in the weblink you sent, I was able to write a programming for bootstrap, but I got the following error

"insufficient observations to compute bootstrap standard errors no results will be saved"

Interestingly, the above error disappears (that means the program I wrote works well) if I use a smaller sample size, say using dataex command to randomly draw 500 observations. However, the error occurs when I use the whole sample. I have tried to use option -nodrop- but it does not work. Do you have any advice for this issue?

Note: I did not provide an example data here because the program works well with the smaller sample as I mentioned above.

The program I wrote is as follows:

capture program drop example
program define example, rclass    
    reg y1 z1 x1, robust
    capture drop yhat1
    predict yhat1
    
    reg y2 z2 x2, robust
    capture drop yhat2
    predict yhat2
    
    reg y yhat1 yhat2 x1 x2, robust
    return scalar y1 = _b[yhat1]
    return scalar y2 = _b[yhat2]
    
end

bootstrap y1coef = (r(y1)) y2coef = (r(y2)), seed(123) reps(50) nodrop: example

[Fei Wang]

Duong, I didn't see problems in your code. There might be issues in your whole sample and the subsample you drew may happen to avoid the issues. You may try running the bootstrap code with the first 500 obs, then the first 600, and so on, to see when the error message pops up.