Ultimately, geolocation based on Range-Doppler will be as accurate as the exactness of the height data you provide. S1TBX Terrain Correction and Ellipsoid Correction use two different sets of height data or models. In no-DEM areas, TC uses a gravitational model (geoid); EC uses the annotated heights (which for S1 are referenced to sea level) and, by design or by omission, no gravitational model. Which operation (TC or EC) is more accurate for a given no-DEM image will depend on which of the height models better represents the true height of the scene. For ocean-only images, TC will be more accurate, since the geoid is a good representation of the true height of the sea surface (EC will be as accurate as TC in the locations where geoid height = 0). I don’t think it is possible to predict which method will be more accurate in no-DEM land areas: it will depend on true height, geoid height and annotated height (is it 0 for no-DEM areas?).
The geolocation error you are seeing when you EC the Miami scene is because in that part of the world the geoid height is around -30m. EC does not take this height into account, and the result is a shift in the range direction. I’d expect S1TBX TC to be more accurate in that image even if no DEM were available.
Edit: spelling (is -> if)