The HPSEVERITY Procedure

PROC HPSEVERITY enables you to estimate model parameters by minimizing your own objective function. This example illustrates how you can use PROC HPSEVERITY to implement the Cramér-von Mises estimator. Let denote the estimate of CDF at for a distribution with parameters , and let denote the empirical estimate of CDF (EDF) at that is computed from a sample , . Then, the Cramér-von Mises estimator of the parameters is defined as

This estimator belongs to the class of minimum distance estimators. It attempts to estimate the parameters such that the squared distance between the CDF and EDF estimates is minimized.

The following PROC HPSEVERITY step uses the Cramér-von Mises estimator to fit four candidate distribution models, including the LOGNGPD mixed-tail distribution model that was defined in Defining a Model for Mixed-Tail Distributions. The input sample is the same as is used in that example.

/*--- Set the search path for functions defined with PROC FCMP ---*/
options cmplib=(work.sevexmpl);

/*-------- Fit LOGNGPD model with PROC HPSEVERITY by using -------
  -------- the Cramer-von Mises minimum distance estimator -------*/
proc hpseverity data=testmixdist obj=cvmobj print=all;
   loss y;
   dist logngpd burr logn gpd;

   * Cramer-von Mises estimator (minimizes the distance *
   * between parametric and nonparametric estimates)    *;
   cvmobj = _cdf_(y);
   cvmobj = (cvmobj -_edf_(y))**2;
run;

The OBJ= option in the PROC HPSEVERITY statement specifies that the objective function cvmobj should be minimized. The programming statements compute the contribution of each observation in the input data set to the objective function cvmobj. The use of keyword functions _CDF_ and _EDF_ makes the program applicable to all the distributions.

Some of the key results prepared by PROC HPSEVERITY are shown in Output 9.7.1. The “Model Selection” table indicates that all models converged. When you specify a custom objective function, the default selection criterion is the value of the custom objective function. The “All Fit Statistics” table indicates that LOGNGPD is the best distribution according to all the statistics of fit. Comparing the fit statistics of Output 9.7.1 with those of Output 9.3.1 indicates that the use of the Cramér-von Mises estimator has resulted in smaller values for all the EDF-based statistics of fit for all the models, which is expected from a minimum distance estimator.