Gretl 2020a windows 64bit.
Consider the following output from mle estimation (pay attention to the
Gradient values and the NA std errors):
<<
--- FINAL VALUES:
loglikelihood = -1241.43371416 (steplength = 6.5536e-012)
Parameters: 1.3798 1.0173 0.96596 3.7418e-012
Gradients: 2.4007 6.7772 -5.3557 -0.89559 (norm
1.96e+000)
Tolerance = 1.81899e-012
Function evaluations: 273
Evaluations of gradient: 39
Model 2: ML, using observations 1-1000
logl = check ? obsll :NA
Standard errors based on Outer Products matrix
estimate std. error z p-value
---------------------------------------------------
const 1.37979 NA NA NA
X1 1.01730 NA NA NA
sv 0.965960 NA NA NA
su 3.74177e-012 NA NA NA
Warning: norm of gradient = 1.96088
>
I have a question, and a suspicion about what the answer might be.
My question is: since the Std errors are based on the Outer Products
matrix, why are they not computed/presented ? Never mind whether this is
a successful estimation attempt -it is not. My question is "mechanical".
I would understand "NA" if I was asking for std errors based on the
inverted Hessian. But where is the problem in computing std errors based
on the gradient?
My suspicion: maybe gretl in all cases dutifully computes the full
variance expression, [G= likelihood gradient], namely Variance = E(GG')
- E(G)E(G)' (using sample means), i.e. it does not a priori treat as
given that E(G)=0. Then, if it so happens that E(G) >> 0 (as is my
case), the full expression may become negative, in which case std errors
are reported as NA since they are the square root of the variance.
Somebody please?
--
Alecos Papadopoulos PhD
Athens University of Economics and Business
web:
alecospapadopoulos.wordpress.com/
skype:alecos.papadopoulos