Nevo, Dorit
>
> I'm not sure I'm going in the right direction so I'm
> pasting the full
> program below. I am trying to calculate the mean and
> standard error of
> sigma. I need quite a few calculations along the was so I tried to
> generalize on the basic monte carlo example that I found in
> the reference
> book and online.
>
> By the time I create the matrix Fc should already be
> calculated. It works
> fine without the matrix command so I'm not sure why the
> problem occurred.
>
> Dorit
>
>
> program define aes_monte
> version 7.0
> if "`1'" == "?" {
> global S_1" mean var"
> exit
> }
> drop _all
> set obs 3000
> gen B0 = -.0353521
> gen R = .5022559 + .0403609*invnorm(uniform())
> gen Dc = -.3476753 + .0700479*invnorm(uniform())
> gen Dk = .3394081 + .0281788*invnorm(uniform())
> gen Bcc = -.0106232 + .0030351*invnorm(uniform())
> gen Bkk = .0782757 + .0023668*invnorm(uniform())
> gen Bll = .0859557 + .0030879*invnorm(uniform())
> gen Bck = .0217035 + .002353*invnorm(uniform())
> gen Bcl = .0066442 + .0040709*invnorm(uniform())
> gen Bkl = -.1821175 + .0040324*invnorm(uniform())
> gen C = 14.86249
> gen lnC = 2.69884
> gen K = 576.6429
> gen lnK = 6.357223
> gen L = 486.6226
> gen lnL = 6.187489
> gen V =exp(B0 - (1/R)*ln(Dc*(C^(-R)) + Dk*(K^(-R)) +
> (1-Dc-Dk)*(L^(-R))) + Bcc*(lnC^2) + Bkk*(lnK^2) +
> Bll*(lnL^2) + Bck*lnC*lnK
> + Bcl*lnC*lnL + Bkl*lnK*lnL)
>
> gen Z = Dc*C^(-R)+Dk*K^(-R)+(1-Dc-Dk)*L^(-R)
> gen Fc = V*(Dc*C^(-R-1)/Z+Bck/C*lnK+Bcl/C*lnL+2*Bcc/C*lnC)
> gen Fk = V*(Dk*K^(-R-1)/Z+Bck/K*lnC+Bkl/K*lnL+2*Bkk/K*lnK)
> gen Fl =
> V*((1-Dc-Dk)*L^(-R-1)/Z+Bcl/L*lnC+Bkl/L*lnK+2*Bll/L*lnL)
> gen Fcc = Fc^2/V - Fc/C +
> V*((-R*Dc*C^(-R-2))/Z+(Dc^2*C^(-2*R-2)*R)/Z^2+2*Bcc/C^2)
> gen Fkk = Fk^2/V - Fk/K +
> V*((-R*Dk*K^(-R-2))/Z+(Dk^2*K^(-2*R-2)*R)/Z^2+2*Bkk/K^2)
> gen Fll = Fl^2/V - Fl/L +
> V*((-R*(1-Dc-Dk)*L^(-R-2))/Z+((1-Dc-Dk)^2*L^(-2*R-2)*R)/Z^2+
> 2*Bll/L^2)
> gen Fck = (Fc*Fk)/V +
> V*((R*Dc*Dk*C^(-R-1)*K^(-R-1)/Z^2)+Bck/(C*K))
> gen Fcl = (Fc*Fl)/V +
> V*((R*Dc*(1-Dc-Dk)*C^(-R-1)*L^(-R-1)/Z^2)+Bcl/(C*L))
> gen Fkl = (Fk*Fl)/V +
> V*((R*Dk*(1-Dc-Dk)*K^(-R-1)*L^(-R-1)/Z^2)+Bkl/(K*L))
>
> matrix H = (0,Fc,Fk, Fl\Fc, Fcc,
> Fck,Fcl\Fk,Fck,Fkk,Fkl\Fl,Fcl,Fkl,Fll)
> gen detH = det(H)
> matrix Hck = (0,Fc,Fl\Fk,Fkc,Fkl\Fl,Flc,Fll)
> gen detHck = det(Hck)
>
> gen sigma_ck = ((C*Fc+K*Fk+L*Fl)/(C*K))*(-detHck/detH)
> summarize sigma_ck
> post `1' (r(mean)) (r(Var))
> end
The bottom line here is that det(H) is a matrix
function yielding one number. Putting the result
in a variable with 3000 obs will just give you 3000 copies
of the same number. Therefore, as I understand
it, you may need, within this file, to loop over observations,
to extract 16 numbers from each observation and get the determinant
in each case.
However, there is a yet wider context -- this is being called
by something else, presumably -simul-.
Despite the fearsome singularity(*) of these calculations,
it looks as if everything is done row-wise:
that is, should all these done as scalar operations
and the 3000 replications conducted via -simul-?
Nick
[email protected]
(*) pun
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
