DEEPANKAR BASU <basu.15@osu.edu> is using -ml init- to set initial values for
-ml-, but is concerned that -ml max, trace- shows that some of the parameter
values are different at iteration 0.
Here is a portion of the log sent by DEEPANKAR:
> . #delimit cr
> delimiter now cr
>
> . ml init /four=1.0
>
> . ml init /five=0.32
>
> . ml init /seven=0.85
>
> . ml init /eight=1.42
>
> . ml init /ten=0.38
>
> . ml init /eleven=0.59
>
> . ml init /twelve=0.86
>
> . ml max, trace
>
> initial: log likelihood = -72821.591
> rescaling entire vector +.
> rescale: log likelihood = -72819.04
> rescaling equations .......+.+++++++++++++++++++++++...++++++++++++++++++++++++++.
> sign reverse ++++++++++++++++++++++++++.....++++++++++++++++++++++++.++++++++++++++++++++++..+...
> rescaling equations .....................+++++++++++++++++++.....
> rescale eq: log likelihood = -72784.475
Notice that -ml- is reporting that it rescaled the coefficient vector, and
then rescaled it again within each equation. This is where -ml- changed some
of DEEPANKAR's initial values. In this case, -ml max- is calling -ml search-,
which is doing the rescaling. The effect of rescaling will result in a
likelihood value that is greater or equal to the original one using the
original initial values.
DEEPANKAR can use the -search(norescale)- option of -ml max- to turn rescaling
off.
-----
DEEPANKAR also mentioned that -ml- was having a difficult time trying to
produce feasible starting values using the data and likelihood evaluator
DEEPANKAR is working with.
Keep in mind that log(0) == . and 1/0 == ., and -ml- may have a difficult time
when the default starting values (or ones provided by users) result in a
missing value for the log-likelihood.
Transformations that remap the entire real line to the set of feasible values
of a parameter {such as (0,1), (-1,1), or positive values} will enable -ml- to
find feasible starting values.
Thus you might want to change your likelihood evaluator to assume one or more
of your constant-only equations are coming from a transformed metric that
allows the entire real line {such as logit(), atanh(), or ln()}, then your
likelihood evaluator can use the inverse transformation {such as invlogit(),
tanh(), or exp(), respectively} to transform them back to the original metric
to compute the log-likelihood value.
--Jeff
jpitblado@stata.com
*
* 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/