In an influential recent paper, Harvey et al (2015) derive an upper limit to the self-interaction cross section of dark matter (σ_DM/m<0.47 cm² g^-1 at 95% confidence) by averaging the dark matter-galaxy offsets in a sample of merging galaxy clusters. Using much more comprehensive data on the same clusters, we identify several substantial errors in their offset measurements. Correcting these errors relaxes the upper limit on σ_DM/m to ≲2 cm² g^-1, following the Harvey et al prescription for relating offsets to cross sections in a simple solid body scattering model. Furthermore, many clusters in the sample violate the assumptions behind this prescription, so even this revised upper limit should be used with caution. Although this particular sample does not tightly constrain self-interacting dark matter models when analyzed this way, we discuss how merger ensembles may be used more effectively in the future. We conclude that errors inherent in using single-band imaging to identify mass and light peaks do not necessarily average out in a sample of this size, particularly when a handful of substructures constitute a majority of the weight in the ensemble.