A number of people have communicated with me or the statalist about my
-probitiv- and -tobitiv- ado files. in a number of cases, people seem
to be interested in endogenous right hand side variables that are
binary, i.e. for -probitiv- the structural equation of interest is

y1* = a1 y2 + X1 b1 + e1,

y1 = 1 if y1* > 0 
   = 0 otherwise 

[the only dif for tobitiv is that y1 = y1* if y1*>0 ]

with

y2 = 1 if y2* > 0
   = 0, otherwise

where

y2* = a2 y1 + X2 b2 + Z d + e2

so that the instruments for y2*, were it observed, are in the vector
Z.

the probitiv and tobitiv commands as written do allow an option to run
a reduced form equation for y2 as a probit.  these commands then
insert predicted probs, i.e. the estimated Pr(y2 = 0 | X1, X2, Z), in
the right hand side of the probit/tobit equation for y1*.  

note, however, that the right hand side of the y1* equation depends on
the dummy variable y2, not the predicted prob that this dummy is
1. using the predicted probs, as the code currently does, will not
generally yield consistent estimates. to see this, rewrite

y1*	= a1 (y2hat + y2 - y2hat) + X1 b1 + e1
	= a1 y2hat + X1 b1 + a1(y2 - y2hat) + e1
	= a1 y2hat + X1 b1 + v

where

v = a1(y2 - y2hat) + e1

if the structural equations are correctly specified, e1 ~ N(0,s^2).

however, (y2 - y2hat) cannot be normally distributed (hence its
product with the scalar a1 can't be either), so v will not generally
have a normal distribution.

the moral of the story is that while probitiv and tobitiv will run
just fine with the "probit" stage1 option, the resulting estimates
have no asymptotically meaningful interpretation.  

the correct approach in such cases is to use FIML (e.g. bivariate
probit when y1 is binary).......