[R] LME: internal workings of QR factorization --repost

[Izmirlian, Grant (NIH/NCI) [E]]

Hi:

I've been reading "Computational Methods for Multilevel Modeling" by Pinheiro and Bates, 
the idea of embedding the technique in my own c-level code. The basic idea is to rewrite 
the joint density in a form to mimic a single least squares problem conditional upon the 
variance parameters.  The paper is fairly clear except that some important level of detail 
is missing. For instance, when we first meet Q_(i):

/                    \                  /                                 \
| Z_i     X_i   y_i  |                  | R_11(i)     R_10(i)     c_1(i)  |
|                    | =         Q_(i)  |                                 |
| Delta   0     0    |                  |   0         R_00(i)     c_0(i)  |
\                    /                  \                                 /

the text indicates that the Q-R factorization is limited to the first q columns of the 
augmented matrix on the left.  If one plunks the first q columns of the augmented matrix 
on the left into a qr factorization, one obtains an orthogonal matrix Q that is (n_i + q) x q 
and a nonsingular upper triangular matrix R that is q x q.  While the text describes R as a 
nonsingular upper triangular q x q, the matrix Q_(i) is described as a square (n_i + q) x (n_i + q) 
orthogonal matrix.  The remaining columns in the matrix to the right are defined by applying 
transpose(Q_(i)) to both sides.  The question is how to augment my Q which is orthogonal (n_i + q) x q  
with the missing (n_i + q) x n_i portion producing the orthogonal square matrix mentioned in the text?  
I tried appending the n_i x n_i identity matrix to the block diagonal, but this doesn't work as the 
resulting likelihood is insensitive to the variance parameters.

Grant Izmirlian
NCI