3dDespike does detrend the data time series in order to detect spikes, but it then re-trends them afterwards. The purpose of re-trending is to produce a dataset that is as much like the dataset that would have been acquired had the spike noise not happened. The transformation into "deviations from some mean" is only an intermediate step in the process. The output is not so transformed.
Also, the "-1zscore" transformation of 3dmerge does not make sense to apply to EPI time series data, filtered/despiked or not. This transformation is to be applied to derived statistical dataset values -- e.g., to a brick of F-statistics from 3dDeconvolve.
bob cox