This section applies to actions in the following action sets: phreg, quantreg, and regression.
When the method subparameter value is BACKWARD, then the backward elimination technique starts from the full model, which includes all independent effects, and deletes effects one by one until a stopping condition is satisfied. At each step, the effect that shows the smallest contribution to the model is deleted.
In the traditional implementation of backward selection, the statistic that is used to determine whether to drop an effect is significance level. At any step, the least significant predictor is dropped and the process continues until all effects that remain in the model are significant at a specified stay significance level (slStay).
Just as with forward selection, you can use the select subparameter to change the criterion that is used to assess effect contributions. You can also specify a stopping criterion in the stop subparameter and use a choose subparameter to provide a criterion for selecting among the sequence of models produced. For more information, see the discussion in the section Forward Selection.
The following CASL language statement removes effects that at each step produce the largest value of the Schwarz Bayesian information criterion (SBC) statistic and stops at the step where removing any effect increases the SBC statistic:
selection={method='backward',stopHorizon=1}
The following CASL language statement bases removal of effects on the corrected Akaike’s information criterion (AICC) and stops at the first step whose AICC value is smaller than the AICC value at the next three steps:
selection={method='backward',select='AICC'}
The following CASL language statement bases removal of effects on significance level and stops when all candidate effects for removal at a step are significant at the default stay significance level of 0.05:
selection={method='backward',select='SL'}
The following CASL language statement bases removal of effects on significance level and stops when all effects in the model are significant at the 0.1 level. Finally, from the sequence of models generated, the chosen model is the one that produces the smallest average square error when scored on the validation data:
selection={method='backward',select='SL',choose='VALIDATE,sls=0.1}
The following CASL language statement applies in logistic regression models the fast backward technique of Lawless and Singhal (1978), a first-order approximation that has greater numerical efficiency than full backward selection:
selection={method='backward',fast=True}
The fast technique fits an initial full logistic model and a reduced model after the candidate effects have been dropped. On the other hand, full backward selection fits a logistic regression model each time an effect is removed from the model.