"Jerry W. Lewis" wrote:

It has been a while since I did the comparisons, but my vague
recollections are that:
....
- The nonzero coefficients from 2003 LINEST are correct for the full
problem, so you might be able (a new [untested] idea) to fit the terms
that gave zero coefficients to the residuals from the original 2003
LINEST fit, to restore the arbitrarily zeroed coefficients.

Limited testing suggests that when there are coefficients that are exactly
zero, coefficients OTHER THAN THE ASSUMED INTERCEPT TERM (apparently
calculated by subtraction at the center of data) are correct. If you
subtract off the fitted cubic and linear terms from the OP's y-data and then
fit a quadratic (with intercept) without a linear term, then you get the
correct missing coefficients.

Jerry