There's no direct guideline. In general, every model - strictly speaking - would have to be tested for 'mesh convergence', i.e., the mesh would have to be refined vertically and horizontally until no more changes in the result are observed. Practically, this is seldom done - maybe too seldom.
On the other hand, in specific configurations mesh convergence does not exist. For example, refining the mesh around a pumping well placed on a single node, the calculated head at the well node will decrease with mesh refinement, no matter how small the elements become. In this case, there is something like an equivalent well radius, i.e., one specific element size that fits to the real well radius.
