st: peculiar issue generating truncated normals

[Patrick Roland <patrick.rolande@gmail.com>]

I've been trying two ways of generating truncated multivariate normals
in mata, which give different results. I'd be interested if anyone
could shed light on why this might be. In the code I generate
bivariate normal draws with variance c*c', where c = (1,0\1,2) ,
truncated above at (-2,-2).

In the first case, I use simple accept reject sampling. In the second,
I sequentially draw truncated normals (this is explained here, for
example: http://www.hss.caltech.edu/~mshum/gradio/ghk_desc.pdf).

The means are consistently different, across many different seeds. The
first variable has a mean of around -2.41 with accept reject and -2.37
with sequential sampling. I'm running Stata SE 11.2. This seems
peculiar to me - any ideas?

mata
u = (-2,-2)
trials = 100000
c = (1,0\1,2)
GHK = J(trials,2,0)
for(i=1;i<=trials;i++){		
r1 = invnormal(runiform(1,1)*normal(u[1]/c[1,1]))
r2 = invnormal(runiform(1,1)*normal((u[2]-c[2,1]*r1)/c[2,2]))					
GHK[i,.] = (c*(r1\r2))'
}

i=0
AR = J(trials,2,0)
while(i<trials){
	t = c*rnormal(2,1,0,1)
	if(t'<u){
		i=i+1	
		AR[i,.] = t'
	}	
}

mean(GHK)
variance(GHK)
mean(AR)
variance(AR)
end
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/