Difference between estimated coefficients

[Ahmad Alsuliman]

Dear Colleagues

I'm running a mediation analysis in which I run two OLS regressions. They are similar but the second one has 1 more additional control variable and I want to compare the estimated coefficient of the main independent variable in the two regression and conduct a t-test for the difference. The following is an example for the code I'm running (assume that x y and z are controls):

regress Risk index x y 

est store reg1

regress Risk index x y z

est store reg2

suest  reg1 reg2

test [reg1_mean]index-[reg2_mean]index=0

My goal is to compare the Risk coefficient in equation 1 and 2 and run a t test for that. However, when I run the last line of my code (i.e. test...etc), STATA reproduces the regressions but with different coefficients compared to the result if I run the regression one by one myself. Moreover, this code will not run if I add the robust option in the regression command.

Is my code correct? why the regression results differes when running the test code?


the following paper did just that in table 3 (panel A) and deduct the coefficients from each other and test if it is significant or not:

Ferris, S. P., Javakhadze, D., & Rajkovic, T. (2017). CEO social capital, risk-taking and corporate policies. Journal of Corporate Finance, 47, 46-71.


Any advice?


Regards,
Ahmad

[FernandoRios]

HI Ahmad
Can you share the exact results you obtain when estimating the models by hand compared to when using suest? Perhaps there is something in those outputs that can suggest what is going on.
Also, No need to use "robust", suest will provide robust standard errors.
HTH

[Maarten Buis]

You claim that the results are different, but you don't show them, so it is hard for us to comment on that. We need the exact code you used and the exact output.

The hypothesis of interest seems to be that index does not influence Risk via the variable z. So why not build a path model and test that hypothesis?

sysuse nlsw88, clear
gen lnwage = ln(wage)
sem (south -> union lnwage ) ///
(union -> lnwage )
estat teffects, nototal nodirect

[Ahmad Alsuliman]

thank you, Fernando and Maarten

I apologize, I must have done something wrong during all my trials except the last one before I share the results with you!

it is indeed correct and there is no difference in the estimations


@Maarten

my hypothesis is that index influence Risk through variable z. I confess I do not know about the path model but thank you for suggesting it. would you kindly elaborate more about the code you shared and the idea?