The removal of the Galactic and extragalactic foregrounds remains a major challenge for those wishing to make a detection of the Epoch of Reionization (EoR) 21 cm signal. Multiple methods of modelling these foregrounds with varying levels of assumption have been trialled and shown promising recoveries on simulated data. Recently however there has been increased discussion of using the expected shape of the foregrounds in Fourier space to define an EoR window free of foreground contamination. By carrying out analysis within this window only, one can avoid the foregrounds and any statistical bias they might introduce by instead removing these foregrounds. In this paper, we discuss the advantages and disadvantages of both foreground removal and foreground avoidance. We create a series of simulations with noise levels in line with both current and future experiments and compare the recovered statistical cosmological signal from foreground avoidance and a simplified, frequency independent foreground removal model. We find that for current generation experiments, while foreground avoidance enables a better recovery at kperp > 0.6 Mpc-1, foreground removal is able to recover significantly more signal at small klos for both current and future experiments. We also relax the assumption that the foregrounds are smooth. For line-of-sight variations only, foreground removal is able to make a good signal recovery even at 1 per cent while foreground avoidance is compromised significantly. We find that both methods perform well for foreground models with line-of-sight and spatial variations around 0.1 per cent however at levels larger than this both methods are compromised.