[Pictures to follow] Let us start with a thought experiment (which can be simulated): imagine a constant signal with value $c$. Add a full period of a pure sine with non-zero frequency. If you can remove this harmonic contribution by zeroing out its frequency bin in the Fourier domain, then the resulting inverse Fourier signal will still have mean $c$. So remove "some frequency" does not reduce the mean per se.
Now remember that, in the Fourier domain, the "O-th" frequency ought to be equal (up to a constant scale factor) to the average of the signal. This is consistent with the above. The pure sine period has equally distributed values above and below $0$, so this does not contribute to the signal's average.
Then, imagine that the observation time frame only encloses the first positive half of the "one-period sine", or the second one. Zeroing its frequency will sure remove frequency components, yet parts of the "not-otherwise" compensated sine bumps may remain, and modify the amplitude, both ways. Therefore, if everything has been done right, your experiment is not totally surprising.
Finally: if removing or cancelling frequency terms in an FFT was a generic cure (panacea), I would not have the my job, and this site would be leaner. As suggested by @ZRHan, to zero out FFT coefficients is quite dangerous.
