For passive radar interference cancellation, least squares (LS) filter is a popularly used method, which projects the surveillance signal in a subspace orthogonal to interference subspace. Under the assumption of ideal reference signal reconstruction, LS filter is simple and effective to implement. However, there is a limited understanding of its performance in passive sensing environments with non-negligible multipath and noise in reference signal. This study examines such effects on the performance of interference cancellation by LS filter. Since transmitted signal is beyond access in passive radar, modulation error rate (MER) is introduced to characterise the direct-path signal noise ratio (SNR) of the reference signal. Closed-form expression for cancellation loss ratio (CLR) is derived, which quantifies to what extent the SNR and interference noise ratio of reference signal and surveillance signal should be achieved to guarantee an acceptable SNR loss of interested target after interference cancellation. Moreover, a modified-MER scheme is proposed to cope with the noise residual in the reconstructed reference signal so as to improve cancellation performance. Effectiveness is verified by both simulated and experimental collected passive radar data using digital terrestrial multimedia broadcasting waveform. These results are useful in practical data processing for passive radar sensing.