tl;dr Summary: I would not use a Heckman selection model (Type II Tobit) if the economics of the problem allow me to fit the the Type I Tobit. If that's not the case, I would be wary of doing the Type II Tobit/Heckman Selection D-in-D unless I had an exclusion restriction. If I did not have one, I would prefer simple OLS since the bias may be smaller.
Longer Version:
The Heckman selection model separates the intensive margin decision (whether to participate at all) and the extensive margin decision (how much to participate) into two equations, where the 2 error terms can be correlated. One example is the decision to work and the decision about how many hours to work. Both may be affected by unobserved energy or motivation. Participation ($y_{1}=1)$ happens when the latent variable $y_{1}^*>0$, and then you get to observe $y_{2}=y_{2}^*$ for those who choose to participate. Those who do not participate have missing outcomes $y_{2}$ in the Heckman model (and zeros in the Tobit world).
Variables that determine participation and intensity can be (and should be) somewhat different in the Heckman model. The Type I Tobit is special case when $y_{1}^*=y_{2}^*$, and so the explanatory variables are then identical. In that case, it seems like you should just use the tobit with the D-in-D approach. Below, I show that using the Heckman selection model seems like a bad idea in a Tobit world. It's inconsistent.
If your latent variables are different, you can't use the Tobit, but you should probably use the MLE rather than the 2 stage version since you have access to Stata (where it's pretty easy) and if your data is not too big. This seems to recover the coefficients very well. However, if you don't have an exclusion restriction (where $x$ determines participation, but not intensity), 2 stage won't work and the you would be better off doing OLS with $y2$.
Stata Code:
clear
set obs 10000
gen id=_n
gen TG = mod(_n,2)
expand 2
bys id: gen after =_n
replace after = after - 1
set seed 12345
/* Tobit Model */
gen ystar = .5 + TG + after*TG *3 + rnormal()
bys TG after: sum ystar
gen y1 = cond(ystar>0,1,0) // extensive margin
gen y2 = cond(ystar>0,ystar,0) // intensive margin
tobit y2 i.after##i.TG, ll(0)
replace y2 = cond(ystar>0,ystar,.)
heckman y2 i.after##i.TG, select(y1 = i.after##i.TG) twostep // 2 Stage Version
heckman y2 i.after##i.TG, select(i.after##i.TG) nolog // MLE Version
drop y1 y2 ystar
/* Heckman Selection Model WITH Exclusion Restriction */
matrix C = (1, .25 \ .25, 1)
drawnorm eps1 eps2, n(`_N') corr(C)
gen x = rnormal()
gen y1star = .5*TG + after*TG *2 + 1/4 *x + eps1
gen y2star = .5 + TG + after*TG *3 +0 *x + eps2
gen y1 = cond(y1star>0,1,0) // extensive margin
gen y2 = cond(y1star>0,y2star,.) // intensive margin
heckman y2 i.after##i.TG, select(y1 = i.after##i.TG x) twostep // 2 Stage Version
heckman y2 i.after##i.TG, select(y1 = i.after##i.TG x) nolog // MLE Version
reg y2 i.after##i.TG x // OLS
drop y1* y2* eps* x
/* Heckman Selection Model WITHOUT Exclusion Restriction */
matrix C = (1, .25 \ .25, 1)
drawnorm eps1 eps2, n(`_N') corr(C)
gen x = rnormal()
gen y1star = .5*TG + after*TG *2 + 0 *x + eps1
gen y2star = .5 + TG + after*TG *3 + 0 *x + eps2
gen y1 = cond(y1star>0,1,0) // extensive margin
gen y2 = cond(y1star>0,y2star,.) // intensive margin
heckman y2 i.after##i.TG, select(y1 = i.after##i.TG x) twostep // 2 Stage Version
heckman y2 i.after##i.TG, select(y1 = i.after##i.TG x) nolog // MLE Version
reg y2 i.after##i.TG x // OLS
drop y1* y2* eps* x
Best Answer
Error terms can be correlated because they represent a common shock. For example, in a model of wage determination, where there is selection into the labour market (work or not), a negative demand shock for labour can affect how much individuals desire to work, and their wages.
This scenario can exist even if regressors are exogenous (which is what you mean by explanatory variables being independent of error terms).