The SEVERITY procedure enables you to estimate the effects of regressor (exogenous) variables while fitting a distribution if the distribution has a scale parameter or a log-transformed scale parameter.
Let () denote the regressor variables. Let denote the regression parameter that corresponds to the regressor . If regression effects are not specified, then the model for the response variable is of the form
|
where is the distribution of with parameters . This model is typically referred to as the error model. The regression effects are modeled by extending the error model to the following form:
|
Under this model, the distribution of is valid and belongs to the same parametric family as if and only if has a scale parameter. Let denote the scale parameter and denote the set of nonscale distribution parameters of . Then the model can be rewritten as
|
such that is affected by the regressors as
|
where is the base value of the scale parameter. Thus, the regression model consists of the following parameters: , , and .
Given this form of the model, distributions without a scale parameter cannot be considered when regression effects are to be modeled. If a distribution does not have a direct scale parameter, then PROC SEVERITY accepts it only if it has a log-transformed scale parameter — that is, if it has a parameter .
The regression parameters are initialized either by using the values you specify or by the default method.
If you provide initial values for the regression parameters, then you must provide valid, nonmissing initial values for and parameters for all .
You can specify the initial value for using either the INEST= data set or the INIT= option in the DIST statement. If the distribution has a direct scale parameter (no transformation), then the initial value for the first parameter of the distribution is used as an initial value for . If the distribution has a log-transformed scale parameter, then the initial value for the first parameter of the distribution is used as an initial value for .
You can use only the INEST= data set to specify the initial values for . The INEST= data set must contain nonmissing initial values for all the regressors specified in the SCALEMODEL statement. The only missing value allowed is the special missing value .R, which indicates that the regressor is linearly dependent on other regressors. If you specify .R for a regressor for one distribution in a BY group, you must specify it so for all the distributions in that BY group.
If you do not specify valid initial values for or parameters for all , then PROC SEVERITY initializes those parameters using the following method:
Let a random variable be distributed as , where is the scale parameter. By definition of the scale parameter, a random variable is distributed as such that . Given a random error term that is generated from a distribution , a value from the distribution of can be generated as
|
Taking the logarithm of both sides and using the relationship of with the regressors yields:
|
PROC SEVERITY makes use of the preceding relationship to initialize parameters of a regression model with distribution dist as follows:
The following linear regression problem is solved to obtain initial estimates of and :
|
The estimates of in the solution of this regression problem are used to initialize the respective regression parameters of the model. The estimate of is later used to initialize the value of .
The results of this regression are also used to detect whether any regressors are linearly dependent on the other regressors. If any such regressors are found, then a warning is written to the SAS log and the corresponding regressor is eliminated from further analysis. The estimates for linearly dependent regressors are denoted by a special missing value of .R in the OUTEST= data set and in any displayed output.
Let denote the initial value of the scale parameter.
If the distribution model of dist does not contain the dist_PARMINIT subroutine, then and all the nonscale distribution parameters are initialized to the default value of 0.001.
However, it is strongly recommended that each distribution’s model contain the dist_PARMINIT subroutine. See the section Defining a Distribution Model with the FCMP Procedure for more information. If that subroutine is defined, then is initialized as follows:
Each input value of the response variable is transformed to its scale-normalized version as
|
where denotes the value of th regressor in the th input observation. These values are used to compute the input arguments for the dist_PARMINIT subroutine. The values that are computed by the subroutine for nonscale parameters are used as their respective initial values. If the distribution has an untransformed scale parameter, then is set to the value of the scale parameter that is computed by the subroutine. If the distribution has a log-transformed scale parameter , then is computed as , where is the value of computed by the subroutine.
The value of is initialized as
|
When you request estimates to be written to the output (either ODS displayed output or in the OUTEST= data set), the estimate of the base value of the first distribution parameter is reported. If the first parameter is the log-transformed scale parameter, then the estimate of is reported; otherwise, the estimate of is reported. The transform of the first parameter of a distribution dist is controlled by the dist_SCALETRANSFORM function that is defined for it.
When regression effects are estimated, the estimate of the scale parameter depends on the values of the regressors and the estimates of the regression parameters. This dependency results in a potentially different distribution for each observation. To make estimates of the cumulative distribution function (CDF) and probability density function (PDF) comparable across distributions and comparable to the empirical distribution function (EDF), PROC SEVERITY reports the CDF and PDF estimates from a representative distribution. The representative distribution is a mixture of a certain number of distributions, where each distribution differs only in the value of the scale parameter. You can specify the number of distributions in the mixture and how their scale values are chosen by using the DFMIXTURE= option in the SCALEMODEL statement.
Let denote the number of observations used for estimation, denote the number of components in the mixture distribution, denote the scale parameter of the th mixture component, and denote the weight associated with th mixture component.
Let and denote the PDF and CDF, respectively, of the th component distribution, where denotes the set of estimates of all parameters of the distribution other than the scale parameter. Then, the PDF and CDF estimates, and , respectively, of the mixture distribution at are computed as follows:
|
|
|
|
where is the normalization factor ().
The CDF estimates reported in OUTCDF= data set, plotted in CDF plots, and used for computing the EDF-based statistics of fit are the values. The PDF estimates plotted in PDF plots are the values.
The scale values for the mixture components are derived from the set () of scale values, where denotes the estimate of the scale parameter due to observation . It is computed as
|
where is an estimate of the base value of the scale parameter, are the estimates of regression coefficients, and is the value of regressor in observation .
Let denote the weight of observation . If the WEIGHT statement is specified, then it is equal to the value of the specified weight variable for the corresponding observation in the DATA= data set; otherwise, it is set to 1.
You can specify one of the following method-names in the DFMIXTURE= option in the SCALEMODEL statement to specify the method of choosing and the corresponding and values:
In this method, there are as many mixture components as the number of observations that are used for estimation. In other words, , , and (). This is the slowest method, because it requires computations to compute the mixture CDF or the mixture PDF of one observation. For observations, the computational complexity in terms of number of PDF or CDF evaluations is . Even for moderately large values of , the time taken to compute the mixture CDF and PDF can significantly exceed the time taken to estimate the model parameters. So, it is recommended that you use this method only for small data sets.
In this method, the mixture contains only one distribution, whose scale value is the mean of the scale values that are implied by all the observations. In other words, is computed as
|
where is the total weight ().
This method is the fastest because it requires only one CDF or PDF evaluation per observation. The computational complexity is for observations.
If you do not specify the DFMIXTURE= option in the SCALEMODEL statement, then this is the default method.
In this method, a certain number of quantiles are chosen from the set of all scale values. If you specify a value of for the K= option when specifying this method, then and are set to be the -quantiles from the set (). The weight of each of the components () is assumed to be 1 for this method.
The default value of is 2, which implies a one-point mixture with a distribution whose scale value is equal to the median scale value.
For this method, PROC SEVERITY needs to sort the scale values in the set , which requires computations. Then, computing mixture estimate of one observation requires CDF or PDF evaluations. Hence, the computational complexity of this method is for computing a mixture PDF or CDF of observations. For , it is significantly faster than the FULL method.
In this method, a uniform random sample of observations is chosen and the mixture contains the distributions that are implied by those observations. If you specify a value of for the K= option when specifying this method, then the size of the sample is . Hence, . If denotes the index of th observation in the sample (), such that , then the scale of th component distribution in the mixture is and the weight associated with it is .
You can also specify the seed to be used for generating the random sample by using the SEED= option for this method. The same sample of observations is used for all models.
Computing a mixture estimate of one observation requires CDF or PDF evaluations. Hence, the computational complexity of this method is for computing a mixture PDF or CDF of observations. For , it is significantly faster than the FULL method.