Re: [Gretl-users] series of estimated log-likelihood values - Clarification for accessor $llt
On Sat, 13 Jun 2015, Alecos Papadopoulos wrote:
In the Help we read, as regards the accessor $llt
<<
$llt
Output: series
For selected models estimated via Maximum Likelihood, returns the series of
per-observation log-likelihood values.
At present this is supported only for binary logit and probit, tobit and
heckit.
>>
But what if I estimate a model using the mle command in a hand-made a
script? Will the accessor work?
It appears not:
After the "end mle" line I typed
series llobs = $llt
and I got "The statistic you requested is not available". But maybe I have
not incorporated it correctly.
Motive: I need the per-observation log-likelihood values to run Vuong's
(1989) LR tests to select between non-nested models.
As the help says, the $llt accessor isn't supported after an mle block.
However, there's a very easy solution. If you predefine the series which
contains the loglikelihood during the mle iteration, it will survive the
mle command and be available for further use. For example:
<hansl>
nulldata 100
p = 0.45
series x = uniform() < p
series ll = NA
mle ll = ln(p) * x + ln(1-p) * (1-x)
params p
end mle
</hansl>
And the series "ll" will contain what you need.
-------------------------------------------------------
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
-------------------------------------------------------