I'm not familiar with that method, but my understanding from reading some online documentation (e.g.
http://en.wikipedia.org/wiki/Bartlett's_method[
^],
http://math.mit.edu/linearalgebra/ila0403.pdf[
^], etc.) I come to the following conclusion:
- Bartlett's method basicalliy calculates a series of local "gravity points" per choosen segmentation (what you call groups, I guess).
Quote from Bartlett's Method[^]: "[...] It provides a way to reduce the variance of the periodogram in exchange for a reduction of resolution, compared to standard periodograms [...]" - Bartlett's method reduces the number of data points by this "gravity point" calculation - it does not calculate any regression curve.
- To calculate the regression from the calculated "gravitiy points": Having two segments only does not take in consideration that there is noise (or "errors" as you name them) for the given nature of the problem. Taking three segments respects that fact.
- Not sure if adding more segments (more "gravity points") give a better linear regression. I guess not, assuming each segment contains "sufficient" data points to calculate the "gravity point" of each segment.
For practical purposes, I think (guess) it is sufficient to make three segments only.
Cheers
Andi