Why tweaking filter makes S11 change while S21 is constant?

I can't understand how S11 can be constant, while S21 changes (in software simulation).
For example, same filter with small adjustment in configuration:
configuration1: S21=-2.5dB, S11=-11dB
configuration2: S21=-2.5dB, S11=-17dB

If i use real measurements with network analyzer, would it be similar? Where that 17dB-11dB=6dB comes from if S21 is almost constant?

You are right: this is only possible if the losses change (dissipation or radiation).
For constant losses in the filter, |S11|^2 + |S21|^2 = const

Is related to loaded-Q tuning of the resonators part of the filter.
This is the reason that sometime a multi-pole filter can have less inband insertion loss (S21), than let's say a 3-pole filter.
The resonators part of the multi-pole filter could get higher Q because they are less "loaded" by the input/output impedances.