Estimates in subsample regression vs. full sample regression

[Peter Li]

Here are my specifications:

***************
reg std_yrg_math if black==1
g adj_black_1=_b[_cons]

reg std_yrg_math black asian hisp other white, nocon
g adj_black_2=_b[black] 

reg std_yrg_math black asian hisp other white
g adj_black_3=_b[black] + _b[_cons]

***************
reg std_yrg_math control1 control2 control3 if black==1
g adj_black_4=_b[_cons]

reg std_yrg_math black asian hisp other white control1 control2 control3, nocon
g adj_black_5=_b[black] 

reg std_yrg_math black asian hisp other white control1 control2 control3
g adj_black_6=_b[black] + _b[_cons]

***************
sum adj_black_*

Output:

    Variable |        Obs        Mean    Std. Dev.       Min        Max
-------------+---------------------------------------------------------
 adj_black_1 |    132,661   -.3097087           0  -.3097087  -.3097087
 adj_black_2 |    132,661   -.3097087           0  -.3097087  -.3097087
 adj_black_3 |    132,661   -.3097087           0  -.3097087  -.3097087
 adj_black_4 |    132,661    .0119388           0   .0119388   .0119388
 adj_black_5 |    132,661    .0650491           0   .0650491   .0650491
 adj_black_6 |    132,661    .0650491           0   .0650491   .0650491

I can't figure out what makes the coefficient of the fourth specification (highlighted red) different?
I'm thinking the constant in this case is the subsample (black) mean adjusting for the three controls, isn't it?
Could someone explain this to me?

Thank you very much!

[William Lisowski]

The effect of the three controls may be different in the Black population than it is in the full population.

[Peter Li]

William Lisowski Thank you! What could I do to prove that's the case?

[William Lisowski]

reg std_yrg_math control1 control2 control3 if black==1
reg std_yrg_math control1 control2 control3

and look at the results. You will see that the estimated coefficients are different.

Or

reg std_yrg_math i.black##(c.control1 c.control2 c.control3)

(assuming the controls are continuous variables) and notice that the interaction effect of black and the controls is nonzero.