We propose a decision rule for symbol erasure insertion in a dispersive propagation environment based on Bayesian decision theory. The method is applied to frequency-hop communications with Reed-Solomon coding and error-and-erasures decoding. The proposed algorithm is based on decision feedback to eliminate statistical dependence among the envelope detectors' outputs of an M-ary frequency-shift keying noncoherent demodulator. The performance is analyzed for fading channels with different levels of frequency selectivity and synchronous multiple-access interference. It is shown that the use of memoryless decision rules results in complete inefficiency in such channel conditions, whereas the proposed rule provides large coding gains even for severe selectivity. The performance sensitivity with respect to the estimation of the power delay profile denotes high robustness of the proposed approach, provided that the selectivity is somewhat overestimated. Sensitivity results for error propagation due to decision feedback are also given.