Re: Reviewing covariance-matrix.c and interaction.c

[Jason Stover]

On Thu, Jun 18, 2009 at 08:26:08PM -0700, Ben Pfaff wrote:
> Some of these modifications to glm.q add debug printing code,
> which presumably helps you to better understand what is happening
> inside.  But some of it, I think, changes what the GLM procedure
> does.  If I'm right about that, then is one version or the other
> "more correct"; otherwise, what is the reason for those changes?

I changed glm.q to use it for testing the covariance code,
not for fitting linear models. The final glm.q would use the
covariance and interaction code in the modified version, mixed
with something like the model-fitting code in regression.q.

So which glm.q is "more correct" depends. The modified version is only
"correct" in the sense of using the new covariance/interaction code,
not the linreg model-fitting code.

> (I'm trying to understand some of the background here.  Maybe I
> can review interaction.c and covariance-matrix.c without that
> background, but it would be nice to understand more.)

I will start changing the master copy of glm.q to properly fit
linear models with the new covariance code. That should clarify
its use some.