Backward Elimination

A model building technique where first a model is built including all potential predictor variables. Predictors are then removed from the model one at a time, typically starting with the predictor with the least contribution to the model quality (accuracy). Backward elimination of predictors continues as long as desired model quality does not decrease below a certain value, or when no other variables in the model equations can be removed without significantly decreasing model quality (given some user-defined criterion). In a variation of this method, after removal of a predictor from the equation, all predictors not in the equation are checked for their unique potential contribution to the model quality, and the predictor with the highest contribution may then be entered into the model equation before the next backward elimination step.