Various methods have been proposed to overcome the problem of compensating the acoustic feedback path that negatively impacts the performance of hearing aid devices. However, in most of them feedback path model is assumed to be fixed which is not quite realistic. In this paper, we consider fixed and variable feedback paths and analyze for each case the performance of one of the robust Adaptive Feedback Cancellation (AFC) schemes, i.e. the Prediction Error Method AFC which uses Partitioned Block Frequency-Domain Normalized Least Mean Square (PBFD-NLMS) algorithm. Based on the analysis results we propose varying the step size values for the same adaptive algorithm on the fly by monitoring the misalignment criteria. The experimental results using the proposed method show improvement made on the system performance.