The problem of state-space model reduction and, in particular, the identification of essential states is considered. The method of a state/output energy index is reviewed and shown, in some cases, to yield insufficient information to determine the essential states. The modal indices of Litz are reassessed, exposing their explicit link with controllability. A new method of essential-state selection is proposed. First the dominant modes are identified, and then these are used to indicate which states contain a significant contribution from these dominant modes. The algorithm can be used to select the essential states in either the state vector or the output. The method is demonstrated on a model having the structure of a rolling mill stand model.