Predictive deconvolution is an effective way to suppress multiple reflections, especially short path multiples, in seismic data. However, the effectiveness of the predictive deconvolution decreases with increasing offset for the shot data. The design and application of the predictive deconvolution filter are based on two significant parameters, prediction lag and operator length, which are generally used as constant along the offset in the application of conventional predictive deconvolution (CPD). In addition to the effect of the wavelet characteristics of the input seismic data, the operator length is related to the performance of the predictive deconvolution filter while the prediction lag controls the temporal resolution of the input trace. If these parameters are determined by considering near offsets on the autocorrelation window, the primary reflections at far offsets will be attenuated. Conversely, if considering far offsets, then the performance of the predictive deconvolution filter reduces in the elimination of the multiples. In this study, to overcome this trade-off between the near and far offsets, these two parameters are used as variable with increasing offset and it is called offset-dependent predictive deconvolution (ODPD). Detailed analyses of this approach were performed on two synthetics including two specific types of short path multiples such as water bottom peg-leg, intrabed and real marine seismic data. It is observed that operator length should be selected long enough to improve the performance of the predictive deconvolution, especially at the nearest offsets. In addition, if the prediction lag is taken longer than second zero crossing of the autocorrelogram, then the reflection amplitudes from particularly deeper layers are better preserved. It is concluded that the use of the offset varying parameters increases the efficiency of the predictive deconvolution filter and preserves reflection amplitudes, leading to a better signal-to-noise ratio (S/N) of the seismic data.