Model reduction based on selected measures of the output equation error