ESTIMATION Subcommand (SURVREG AFT command)
ESTIMATION is an optional subcommand to specify the settings to control the
estimation of the accelerated failure time models and the optional feature selection process.
- SELECTFEATURES
- An optional parameter to control feature selection. FALSE is the setting by default, which suppresses the feature selection process. TRUE will conduct the process.
- PENALTY
- An optional parameter to specify the penalty parameter that controls the regularization process. It must be a single value greater than 0. The default setting is 0.001. The process will be ignored if SELECTFEATURES is FALSE.
- ADMM
- An optional parameter to specify the type of the alternating direction method of multipliers (ADMM). FAST is the setting by default, which will apply the fast ADMM. TRADITIONAL will apply the traditional ADMM algorithm.
- METHOD
- An optional parameter to specify the estimation method. AUTO is the setting by default, which will automatically choose the method based on the sample data set. If specified, it could optionally be followed by an additional parameter THRESHOLD specified within parentheses to denote the threshold of the number of predictors. If specified, it must be a single integer greater than 1. The default value is 1000. NEWTON_RAPHSON will apply the Newton-Raphson's method. L_BFGS will apply the limited-memory BFGS algorithm. If specified, it could optionally be followed by an additional parameter UPDATE specified within parentheses to denote the number of the past updates maintained by the limited_memory BFGS algorithm. If specified, it must be a single integer greater than or equal to 1. The default value is 5.
- MAXLINESEARCH
- An optional parameter to specify the maximum number of the line searches. It must be a single integer belonging to [1, 100]. The default setting is 20.
- MAXITER
- An optional parameter to specify the maximum number of iterations. It must be a single integer belonging to [1, 100]. The default setting is 20.
- MAXSTEPHALVING
- An optional parameter to specify the maximum number of step-halving. It must be a single integer belonging to [1, 20]. The default setting is 5.
- PCONVERGE
- An optional parameter to specify the convergence criteria for the parameter. It must be a single numeric value belonging to [0, 1). The default setting is 0.000001. The value could optionally be followed by a keyword specified in parentheses. The default setting is ABSOLUTE, which will apply the absolute convergence to the inner optimization. RELATIVE will apply the relative convergence to the inner optimization. Note that at least one of the values specified for PCONVERGE, FCONVERGE, and HCONVERGE must be greater than 0.
- FCONVERGE
- An optional parameter to specify the convergence criteria for the objective function. It must be a single numeric value belonging to [0, 1). The default setting is 0, which does not apply the convergence criteria. The value could optionally be followed by a keyword specified in parentheses. The default setting is ABSOLUTE, which will apply the absolute convergence to the inner optimization. RELATIVE will apply the relative convergence to the inner optimization.
- HCONVERGE
- An optional parameter to specify the convergence criteria for the Hessian matrix. It must be a single numeric value belonging to [0, 1). The default setting is 0, which does not apply the convergence criteria. The value could optionally be followed by a keyword specified in parentheses. The default setting is ABSOLUTE, which will apply the absolute convergence to the inner optimization. RELATIVE will apply the relative convergence to the inner optimization.
- OPTIMIZE
- An optional parameter to control the optimization process. BOTH is the setting by default, which will apply both primal and dual residual convergence criterion. PRIMAL will only apply the primal residual convergence criterion. DUAL will only apply the dual residual convergence criterion.
- EPSILON_ABS
- An optional parameter to specify the absolute convergence for the outer iteration process. It must be a single numeric value belonging to (0, 1). The default setting is 0.0001.
- EPSILON_REL
- An optional parameter to specify the relative convergence for the outer iteration process. It must be a single numeric value belonging to (0, 1). The default setting is 0.01.