Logistic Regression Stepping Options
These options enable you to control the criteria for adding and removing fields with the Stepwise, Forwards, Backwards, or Backwards Stepwise estimation methods.
Number of terms in model (Multinomial models only). You can specify the minimum number of terms in the model for Backwards and Backwards Stepwise models and the maximum number of terms for Forwards and Stepwise models. If you specify a minimum value greater than 0, the model will include that many terms, even if some of the terms would have been removed based on statistical criteria. The minimum setting is ignored for Forwards, Stepwise, and Enter models. If you specify a maximum, some terms may be omitted from the model, even though they would have been selected based on statistical criteria. The Specify Maximum setting is ignored for Backwards, Backwards Stepwise, and Enter models.
Entry criterion (Multinomial models only). Select Score to maximize speed of processing. The Likelihood Ratio option may provide somewhat more robust estimates but take longer to compute. The default setting is to use the Score statistic.
Removal criterion. Select Likelihood Ratio for a more robust model. To shorten the time required to build the model, you can try selecting Wald. However, if you have complete or quasi-complete separation in the data (which you can determine by using the Advanced tab on the model nugget), the Wald statistic becomes particularly unreliable and should not be used. The default setting is to use the likelihood-ratio statistic. For binomial models, there is the additional option Conditional. This provides removal testing based on the probability of the likelihood-ratio statistic based on conditional parameter estimates.
Significance thresholds for criteria. This option enables you to specify selection criteria based on the statistical probability (the p value) associated with each field. Fields will be added to the model only if the associated p value is smaller than the Entry value and will be removed only if the p value is larger than the Removal value. The Entry value must be smaller than the Removal value.
Requirements for entry or removal (Multinomial models only). For some applications, it doesn't make mathematical sense to add interaction terms to the model unless the model also contains the lower-order terms for the fields involved in the interaction term. For example, it may not make sense to include A * B in the model unless A and B are also included in the model. These options let you determine how such dependencies are handled during stepwise term selection.
- Hierarchy for discrete effects. Higher-order effects (interactions involving more fields) will enter the model only if all lower-order effects (main effects or interactions involving fewer fields) for the relevant fields are already in the model, and lower-order effects will not be removed if higher-order effects involving the same fields are in the model. This option applies only to categorical fields.
- Hierarchy for all effects. This option works in the same way as the previous option, except that it applies to all input fields.
- Containment for all effects. Effects can be included in the model only if all of the effects contained in the effect are also included in the model. This option is similar to the Hierarchy for all effects option except that continuous fields are treated somewhat differently. For an effect to contain another effect, the contained (lower-order) effect must include all of the continuous fields involved in the containing (higher-order) effect, and the contained effect's categorical fields must be a subset of those in the containing effect. For example, if A and B are categorical fields and X is a continuous field, the term A * B * X contains the terms A * X and B * X.
- None. No relationships are enforced; terms are added to and removed from the model independently.