The number of water molecules which are expected to be experimentally located by protein crystallography was determined by multiple regression analysis on a test set of 873 known protein crystal structures determined at room temperature and on another set of 33 structures determined at low temperature. The dependence of the number of water molecules included in the protein models as a function of a number of significant regressors, such as resolution, fraction of crystal volume occupied by the solvent, number of residues in the asymmetric unit, fraction of apolar protein surface or secondary structure, has been studied. The number of water molecules included in crystallographic models depends primarily on the resolution at which the structure has been solved, while the temperature of the data collection has only marginal influence. On average, at 2.0 A resolution one water molecule per residue is included in the model, while at 1.0 A resolution about 1.6-1.7 are crystallographically located. At 2.0 A resolution the well known rule-of-thumb of 'one water per protein residue' is confirmed, though the number of water molecules experimentally observed is strongly dependent on resolution. The results presented are useful in assessing the quality of a protein crystal structure, in selecting structural results to be compared and in evaluating the expected improvement on the solvent structure when increasing the crystallographic resolution.