invsym(): 3900 unable to allocate real <tmp> after -predict var, remodes-

[Jeph Herrin]

I estimated a 3-level model using -mepoisson-, which converged without errors using intmethod(gh). The error structure is a little bit complex, with two random intercepts at level 3, a covariance between them, and a single random intercept at level 2. The problem was when I attempted to get the predicted values for the random terms:

. predict RE3mode*, remodes
(calculating posterior modes of random effects)
                invsym():  3900  unable to allocate real <tmp>[29308,29308]
    opt__looputil_step():     -  function returned error
          opt__loop_nr():     -  function returned error
             _optimize():     -  function returned error
             _gsem_agh():     -  function returned error
 _gsem_predict__latent():     -  function returned error
  _gsem_predict_latent():     -  function returned error
                 <istmt>:     -  function returned error
r(3900);

Is this purely an out of memory problem? The dataset is 43mb, the computer, running Windows 7 64-bit, has 8gb of RAM. If purely a memory problem, I'll move to another machine with 24gb.

thanks,
Jeph

[Jeff Pitblado (StataCorp)]

A square matrix with 29,308 diagonal elements takes up almost 7gb.
The call to invsym() means that at some point 2 of these monsters
need to be present at the same time.

I believe moving to the 24gb machine should fix it.