ML evaluator crashes if it calls a built-in maximum likelihood function

[David Berger]

Suppose I want to run a maximum likelihood program that calls another maximum likelihood program. (For example: suppose I have two equations and one of them is essentially an ordered logit, which is rather difficult to program from scratch, as numerical approximations make an lf0 estimator take forever.) So I write something like

program define myeval
  args lnfj Xb
  ologit $ML_y1 `Xb'
  tempvar lnfj_part1
  predict `lnfj_part1', pr
  ...
  replace `lnfj' = log(`lnfj_part1')+...
end

This will produce the output:

variable _MLtua1 not found
               st_data():  3500  invalid Stata variable name
     mopt__st_user_lf1():     -  function returned error
    mopt__calluser_lf2():     -  function returned error
      opt__eval_nr_lf1():     -  function returned error
             opt__eval():     -  function returned error
    _optimize_evaluate():     -  function returned error
       _mopt__evaluate():     -  function returned error
   _moptimize_evaluate():     -  function returned error
     _moptimize_search():     -  function returned error
           Mopt_search():     -  function returned error
                 <istmt>:     -  function returned error
r(3500);

I'm pretty sure the reason is that ml state is stored in global variables, which are overwritten when ologit is called. Is this correct? Does this mean that one can never use an ml estimation command within another ml evaluator, or is there some kind of workaround?

[Jeff Pitblado (StataCorp)]

See the help file for ml hold.

I think the myeval program should be modified as follows:

program define myeval
  args lnfj Xb
  local y1 $ML_y1
  ml hold
  ologit `y1' `Xb'
  ml untold
  tempvar lnfj_part1
  predict `lnfj_part1', pr
  ...
  replace `lnfj' = log(`lnfj_part1')+...
end