The deviations of the second-order state-specific effective Hamiltonian method from the strict size consistency are analyzed. Provided that complete or separable model spaces are used, these deviations can be suppressed by a proper choice of nonuniform shifts of energy denominators.