METHOD Subcommand (COXREG command)

METHOD specifies the order of processing and the manner in which the covariates enter the model. If no METHOD subcommand is specified, the default method is ENTER.

• The subcommand keyword METHOD can be omitted.
• You can list all covariates to be used for the method on a variable list. If no variable list is specified, the default is ALL; all covariates named after WITH on the VARIABLES subcommand are used for the method.
• The keyword BY can be used between two variable names to specify an interaction term.
• Variables specified on CATEGORICAL are replaced by sets of contrast variables. The contrast variables associated with a categorical variable are entered or removed from the model together.
• Three keywords are available to specify how the model is to be built:

  ENTER. Forced entry. All variables are entered in a single step. This is the default if the METHOD subcommand is omitted.

  FSTEP. Forward stepwise. The covariates specified on FSTEP are tested for entry into the model one by one based on the significance level of the score statistic. The variable with the smallest significance less than PIN is entered into the model. After each entry, variables that are already in the model are tested for possible removal based on the significance of the Wald statistic, likelihood ratio, or conditional criterion. The variable with the largest probability greater than the specified POUT value is removed and the model is reestimated. Variables in the model are then again evaluated for removal. Once no more variables satisfy the removal criteria, covariates not in the model are evaluated for entry. Model building stops when no more variables meet entry or removal criteria, or when the current model is the same as a previous one.

  BSTEP. Backward stepwise. As a first step, the covariates specified on BSTEP are entered into the model together and are tested for removal one by one. Stepwise removal and entry then follow the same process as described for FSTEP until no more variables meet entry and removal criteria, or when the current model is the same as a previous one.

• Multiple METHOD subcommands are allowed and are processed in the order in which they are specified. Each method starts with the results from the previous method. If BSTEP is used, all eligible variables are entered at the first step. All variables are then eligible for entry and removal unless they have been excluded from the METHOD variable list.
• The statistic used in the test for removal can be specified by an additional keyword in parentheses following FSTEP or BSTEP. If FSTEP or BSTEP is specified by itself, the default is COND.

COND. Conditional statistic. This is the default if FSTEP or BSTEP is specified by itself

WALD. Wald statistic. The removal of a covariate from the model is based on the significance of the Wald statistic.

LR. Likelihood ratio. The removal of a covariate from the model is based on the significance of the change in the log-likelihood. If LR is specified, the model must be reestimated without each of the variables in the model. This can substantially increase computational time. However, the likelihood-ratio statistic is better than the Wald statistic for deciding which variables are to be removed.

Example

COXREG VARIABLES = SURVIVAL WITH GROUP SMOKE DRINK
 /STATUS SURVSTA (1)
 /CATEGORICAL = GROUP SMOKE DRINK 
 /METHOD ENTER GROUP
 /METHOD BSTEP (LR) SMOKE DRINK SMOKE BY DRINK.

• GROUP, SMOKE, and DRINK are specified as covariates and as categorical variables.
• The first METHOD subcommand enters GROUP into the model.
• Variables in the model at the termination of the first METHOD subcommand are included in the model at the beginning of the second METHOD subcommand.
• The second METHOD subcommand adds SMOKE, DRINK, and the interaction of SMOKE with DRINK to the previous model.
• Backward stepwise regression analysis is then done using the likelihood-ratio statistic as the removal criterion. The variable GROUP is not eligible for removal because it was not specified on the BSTEP subcommand.
• The procedure continues until the removal of a variable will result in a decrease in the log-likelihood with a probability smaller than POUT.