No. There is no reason why S11 and S21 should be correlated. Likewise S22 could be all over the place. If it is a "home made" attenuator then there could be a resonance in the resistive input matching path making S11 rubbish but if the transfer element was good then S21 could be ok.  

Also remember that you are dealing with log scale and "large" changes in S11 may be appear negligible in S21 (for pads at least).

An example:  network analyzer with 0 dBm (1 mW) output and measuring a 3 dB pad.  S21 is -3 dB and the received power is 0.5 mW.  S11, for example may jump between -20 and -30 dB which looks large (10 dB jumps), however S11 is only jumping in this case between 0.01 mW and 0.001 mW.  S21 may be varying by 0.01 mW and still look very flat in dB.

So it is possible that your S21 measurement is varying as much in terms of linear power as your S11 measurement, but because of the log scale it may not appear to be varying as much.

Peter

Thinking about the question in a more basic way, there are four S parameters for a two port network. For a linear passive network S21 and S12 are the same. This leaves three degree of freedom (three independent complex variables). At any given spot frequency the network can be represented by either a PI or a T network of complex impedances, by which I mean either a series R-L or a series R-C chain (or these could be made from parallel elements instead). Each impedance can therefore be adjusted independently of the others.

The bumps (unsmoothness) in the S21 will match up with the bumps in the S11 (rtn loss).
If you had a nearly lossless part (air dielectric adapter, with one support wheet of course), your S11 "VSWR plot" would look smooth and close to 1:1 while your S21 loss would look smooth. So you can get smoothness on both if you plot S11 VSWR and S21, but not if you plot S11 rtn loss.

Most items have multiple dips in the S11 rtn loss across the operating band.

kch