On Thu, 5 Feb 2015, Alecos Papadopoulos wrote:
> I have to evaluate the finite-sample performance of the ML estimator in the
> estimation of a 3-parameter distribution, based on simulated samples.
> The "progressive loop" environment which is set for Monte Carlo studies does
> not seem to support the mle command. I ended up using the index loop which
> supports the mle command, while writing the coefficient vector obtained in
> each loop that is grabbed by $coeff, to successive rows of a pre-defined
> zero matrix. So far so good.
Your solution is perfectly ok. However, you can circumvent the above
limitation by wrapping mle into a function, as in
<hansl>
function scalar homemade_probit(series y, series x, scalar *b0, scalar *b1)
catch mle ll = ln(cnorm(z))
series ndx = b0 + b1*x
series z = y ? ndx : -ndx
params b0 b1
end mle --quiet
scalar err = $error > 0
return err
end function
nulldata 50
scalar H = 999
series x = normal()
loop H --progressive --quiet
series y = 1 + x > normal()
scalar b0 = 0
scalar b1 = 1
scalar err = homemade_probit(y, x, &b0, &b1)
print b0 b1 err
endloop
</hansl>
> The problem is the following: the distribution appears difficult to estimate
> in small samples and rather often the convergence criterion for the mle is
> not met (I don't mind that, this is part of the evaluation of the performance
> of the MLE). But when this happens, /the loop stops. /The total number of
> runs of the loop will possibly require some hours to run. If the loop stops
> every time the mle does not converge, I will have to be constantly overseeing
> the process for the duration, in order to re-ignite the loop. It would be
> very convenient if the loop could continue irrespective of whether the mle in
> a particular loop fails to converge.
>
> Is there a way to tell Gretl not to stop the loop, even if the mle fails, to
> write the container matrix with NANs, and to proceed to the next run;
You may want to use the $error accessor for that. The above script gives
you a rough example.
-------------------------------------------------------
Riccardo (Jack) Lucchetti
Dipartimento di Scienze Economiche e Sociali (DiSES)
Università Politecnica delle Marche
(formerly known as Università di Ancona)
r.lucchetti(a)univpm.it
http://www2.econ.univpm.it/servizi/hpp/lucchetti
-------------------------------------------------------
