Removing certain windows from Umbrella Sampling changes PMF curve

GROMACS version: 2020.4
GROMACS modification: Yes/No

Hello, I have a question about choosing the right sampling windows for Umbrella Sampling (US), and the resulting PMF curve:
I performed the US method for a protein-nanobody complex, and obtained these results:

Black curves were sampled with a constant of 1000, there were non-sampled regions in the reaction coordinate, so more starting conformations were used: messy areas are where windows overlap, and blue peaks are from sampling windows run with a higher constant.

Since it looked quite messy, I decided to remove the overlapping windows, selected in red in this picture:

keeping only green and blue curves:

However, even though I keep the sampling windows for most of the coordinate, including all at the beginning and the end, removing certain sampling windows seems to affect the calculation for the binding energy of the complex (starting at the same point of the first removed window):

So, my question is, what is the correct procedure to follow? I mean, from the questions here in the forum and the tutorial in I understood it was correct to try to fill the unsampled regions by choosing intermediate starting frames, and also incrementing the pull force constant, if we were in a high energy area. But, in my attempts to do so, many starting frames were chosen and evaluated to see how well they complete the histogram, and apparently the final result can depend on which frames would have been chosen and which windows are finally added, so how to evaluate which should have been chosen from the beginning, or which can be discarded in a “correct” method?

Thank you for your time