The mean square error of prediction is proposed as a criterion for selecting variables. This criterion utilizes the values of the predictor variables associated with the future observation and the magnitude of the estimated variance. Mean square error is believed to be more meaningful than the commonly used criterion, the residual sum of squares.