The calculation of temperature PC series needs to be distinguished from the calculation of tree ring PC series. Procedures for the handling of missing data in the temperature dataset were described in the MBH98 Corrigendum SI for the first time in July 2004, specifying that linear interpolation was done, but not specifying the handling of missing data at the beginning and end of records.
I show here a fairly close replication of MBH98 results from the temperature dataset archived in July 2004 at the Corrigendum SI. Unlike tree ring PC networks, stepwise methods were almost certainly not used for temperature principal components. Also, centered calculations appear to have been used for temperature PC series, but were not used for tree ring PC series. von Storch et al.  pointed out that MBH98 incorrectly used weight of the cosine of the latitude, rather than the square root of the cosine of the latitude, but did not show the specific effects of the error on MBH98 temperature PC series.
Here I show that the error leads to significant differences in the temperature PC series after the first few series. Link