This would seem to be physically inconsistent, yet I see it often, and I am trying to understand why (e.g. is there a reason that the reanalysis SSTs are not believable?).
Also, you want to use several different datasets whenever possible to see if your results are robust and not just only an "artifact" of a particular dataset. For instance, reviewers of papers that propose a new analysis technique often ask for the same analysis to be applied to several reanalysis datasets for that reason.