[ADMB Users] Discrepancy between MLEs in the .std and .rep files
Russell Millar
r.millar at auckland.ac.nz
Tue Mar 17 17:35:54 PDT 2009
Dear ADMB'ers,
Please note that the error in the .std file can be of reasonable magnitude.
For example, for the estrone data in Chap 9 of Pawitan's book ("In all
likelihood") a one-way random effects ANOVA gives the between person
var comp as 1.395. The ADMB-RE estimate varies according to bounding,
but in one case it was 1.486 (see below). Note that the variance
component is a dependent variable since the init parameter is on the
log scale, so both the estimate and std dev of the var comp are affected.
Dave - what do you recommend to users? My experimenting suggests that
the error is "not-so-much" when the last parameter has no explicit
bounding (see below). I guess another approach would be to incorporate
an additional throw-away parameter (e.g., parameter v makes
contribution of pow(v,2) to objective function) that appears last
in the init parameter list - but this does seem very klugey.
Regards,
Russell
True var comp (this is always reported correctly in .rep file)
=============
Random effects:
Groups Name Variance
person (Intercept) 1.39546
ADMB-RE estimates in .std file
c.f.
init_bounded_number log_sigma_b(-100,100,2) // log(person effect s.e)
10 sigmab2 1.4860e+00 9.5351e-01
c.f
init_bounded_number log_sigma_b(-5,5,2) // log(person effect s.e)
10 sigmab2 1.3999e+00 8.9825e-01
c.f
init_number log_sigma_b(2) // log(person effect s.e)
10 sigmab2 1.3960e+00 8.9580e-01
More information about the Users
mailing list