The bane of PEF estimation

This is the place where I would like to pat myself on the back for having ``solved'' the problem of missing data. Actually, an important practical problem remains. I've been trying to coax younger, more energetic people to think about it. The problem arises when there is too much missing data.

The bane of PEF estimation is too much missing data

Then all the regression equations disappear. The nonlinear methods are particularly bad because if they don't have a good enough starting location, they can and do go crazy. My only suggestion is to begin with a linear PEF estimator. Shrink the PEF and coarsen the mesh in model space until you do have enough equations. Starting from there, hopefully you can refine this crude solution without dropping into a local minimum.

Another important practical problem remains, that of nonstationarity. We'll see the beginnings of the solution to that problem in chapter .