Eric questioned the choice of the truncation parameter, and we presented the work Nic and Jeff had done (
using ridge regression, direct RLS with no infilling, and the nearest - station reconstructions) that all gave nearly identical results.
Not exact matches
RegEM is an iterative approach for estimating the data covariance of an incomplete data set (and imputing missing values in the process)
using what is sometimes referred to as «
ridge regression».
Part of Eric's post is spent on the choice to
use individual
ridge regression (iRidge) instead of TTLS for our main results.
Second, in their main reconstruction, O'Donnell et al. choose to
use a routine from Tapio Schneider's «RegEM» code known as «iridge» (individual
ridge regression).
But when, as with the Antarctica weather station data we
used, there is not only a lot of missing data and «noise» but also greatly time - varying patterns of missingness (which stations have data missing),
ridge regression (both mridge and iridge) can be expected to, and does, perform significantly better than TTLS.