Re: SPSS 20 + 64 bit OS

Posted by David Marso on Feb 10, 2012; 2:12am
URL: http://spssx-discussion.165.s1.nabble.com/SPSS-20-64-bit-OS-tp5166219p5471244.html

Frank,
Looks like you will have a serious problem with independence, clustering/hierarchy etc which GLM cannot easily address.  I suspect the design matrix might be permutable to something close to a block diagonal which would allow you to partition the "super-model" into several smaller submodels. Doesn't seem that this is likely to be an easy model to formulate much or less estimate quickly.  I don't have any references handy, but you should look into sparse matrices, updating formulas for sparse matrix inverses, what have others who have researched this area used?  SPSS doesn't readily lend itself to 'real-time' analysis.
HTH, David
--

drfg2008 wrote

I'm proceeding under the (INSANE) assumption that Frank is trying to analyze *ALL* 4798 variables at the same time in a single GLM ;-)

David,
it's a live-game model with all teams / players as dummys and results (for example goals) in a minutes- / seconds - timeframe. It's (of course) not about checking, what variable is stat. relevant, but to estimate the mean expected value (goal) as precise as possible and as fast as possible for the time-frame of the specific game. Since the teams / players play in different 'leagues' and the combination of plays and leagues is hard to differentiate (at least for a programmer) and not 'closed' (not limited to a closed number of leagues), all teams / players should be in one (super-) model.

Plus, the model should run as fast as possible.

Hence my question of max. Gig.

(And a bivariate Poisson- distribution (Team1 - Team2), by the way. Found a solution?)

Frank