Adaptive multiple subtraction is a critical and challenging procedure for the widely used surface-related multiple elimination techniques. One of the problems encountered in this procedure is that a good result usually is hard to obtain when primaries and multiples have overlap or when the true and predicted multiples have mismatches, such as wavelet difference, time shift, and scalar inconsistency. We propose an adaptive multiple subtraction method based on constrained independent component analysis (CICA). It combines the advantages of two current adaptive multiple subtraction methods: the independent component analysis method and the multidimensional prediction error filters (PEF) method. In CICA, the prediction error obtained by the PEFs, which measures the lateral continuity of the primaries, is adopted as a constraint term of its objective function.