reads in covariance parameter values from a SAS data set. The data set should contain the numerical variable Estimate or numerical variables named Covpi. The GLIMMIX procedure associates the values for Covpi with the ith covariance parameter.

For data sets containing the numerical variable Estimate, the GLIMMIX procedure fixes the ith covariance parameter value at the value of the ith observation in the data set. A missing value indicates not to fix the particular parameter. PROC GLIMMIX performs one likelihood ratio test for the TESTDATA= data set.

For data sets containing numerical variables named Covpi, the procedure performs one likelihood ratio test for each observation in the TESTDATA= data set. You do not have to specify a Covpi variable for every covariance parameter. If the value for the variable is not missing, PROC GLIMMIX fixes the associated covariance parameter in the null model. Consider the following statements:

data TestDataSet;
  input covp1 covp2 covp3;
  datalines;
. 0 .
0 0 .
. 0 0
0 0 0
;

proc glimmix method=mspl;
  class subject x;
  model y = x age x*age;
  random intercept age / sub=subject type=un;
  covtest testdata=TestDataSet;
run;

Because the matrix is a unstructured matrix, the first observation of the TestDataSet corresponds to zeroing the covariance between the random intercept and the random slope. When the reduced model is fit, the variances of the intercept and slope are reestimated. The second observation reduces the model to one with only a random slope in age. The third reduces the model to a random intercept model. The last observation eliminates the matrix altogether.

Note that the tests associated with the first and last set of covariance parameters in TestDataSet can also be obtained by using keywords:

proc glimmix;
  class subject x;
  model y = x age x*age;
  random intercept age / sub=subject type=un;
  covtest DiagG;
  covtest GLM;
run;

supplies a list of values at which to fix the covariance parameters. A missing value in the list indicates that the covariance parameter is not fixed. If the list is shorter than the number of covariance parameters, missing values are assumed for all parameters not specified. The COVTEST statements that test the random intercept and random slope in the previous example are as follows:

proc glimmix;
  class subject x;
  model y = x age x*age;
  random intercept age / sub=subject type=un;
  covtest 0 0;
  covtest . 0 0;
run;

provides a general facility to test linear combinations of covariance parameters. You can specify one or more sets of coefficients. The position of a coefficient in the list corresponds to the position of the parameter in the "Covariance Parameter Estimates" table. The linear combination of covariance parameters that is implied by each set of coefficients is tested against zero. If the list of coefficients is shorter than the number of covariance parameters, a zero coefficient is assumed for the remaining parameters.

For example, in a heterogeneous variance model with four groups, the following statements test the simultaneous hypothesis :

proc glimmix;
   class A;
   model y = a;
   random _residual_ / group=A;
   covtest 'pair-wise homogeneity'
           general  1 -1  0  0,
                    0  0  1 -1;                     
run;

In a repeated measures study with four observations per subject, the COVTEST statement in the following example tests whether the four correlation parameters are identical:

proc glimmix;
   class subject drug time;
   model y = drug time drug*time;
   random _residual_ / sub=subject type=unr;
   covtest 'Homogeneous correlation'
        general  0 0 0 0 1 -1            ,
                 0 0 0 0 1  0 -1         ,
                 0 0 0 0 1  0  0 -1      ,
                 0 0 0 0 1  0  0  0 -1   ,
                 0 0 0 0 1  0  0  0  0 -1;
run;

Notice that the variances (the first four covariance parameters) are allowed to vary. The null model for this test is thus a heterogeneous compound symmetry model.

The degrees of freedom associated with these general linear hypotheses are determined as the rank of the matrix , where is the matrix of coefficients and is the number of covariance parameters. Notice that the coefficients in a row do not have to sum to zero. The following statement tests :

covtest general 1 -3, 0 0 1;

requests confidence limits or bounds for the covariance parameter estimates. These limits are displayed as extra columns in the "Covariance Parameter Estimates" table.

The following suboptions determine the computation of confidence bounds and intervals. See the section Statistical Inference for Covariance Parameters for details about constructing likelihood ratio confidence limits for covariance parameters with PROC GLIMMIX.

If you do not specify any suboptions, the default is to compute two-sided Wald confidence intervals with confidence level .

requests that the p-value of the likelihood ratio test be computed by the classical method. If is the realized value of the test statistic in the likelihood ratio test,

where is the degrees of freedom of the hypothesis.

enables you to supply degrees of freedom for the computation of p-values from chi-square mixtures. The mixture weights are supplied with the WGHT= option. If no weights are specified, an equal weight distribution is assumed. If is the realized value of the test statistic in the likelihood ratio test, PROC GLIMMIX computes the p-value as (Shapiro 1988)

Note that and that mixture weights are scaled to sum to one. If you specify more weights than degrees of freedom in value-list, the rank of the hypothesis (DF column) is substituted for the missing degrees of freedom.

Specifying a single value for value-list without giving mixture weights is equivalent to computing the p-value as

For example, the following statements compute the p-value based on a chi-square distribution with one degree of freedom:

proc glimmix noprofile;
   class A sub;
   model score = A;
   random _residual_ / type=ar(1) subject=sub;
   covtest 'ELR low' 30.62555 0.7133361 / df=1;
run;

The DF column of the COVTEST output will continue to read regardless of the DF= specification, however, because the DF column reflects the rank of the hypothesis and equals the number of constraints imposed on the full model.

displays the estimates of the covariance parameters under the null hypothesis. Specifying the ESTIMATES option in one COVTEST statement has the same effect as specifying the option in every COVTEST statement.

limits the number of iterations when you are refitting the model under the null hypothesis to number iterations. If the null model does not converge before the limit is reached, no p-values are produced.

displays the values of the covariance parameters under the null hypothesis. This option is useful if you supply multiple sets of parameter values with the TESTDATA= option. Specifying the PARMS option in one COVTEST statement has the same effect as specifying the option in every COVTEST statement.

specifies that starting values for the covariance parameters for the null model are obtained by the same mechanism as starting values for the full models. For example, if you do not specify a PARMS statement, the RESTART option computes MIVQUE(0) estimates under the null model (Goodnight 1978b). If you provide starting values with the PARMS statement, the starting values for the null model are obtained by applying restrictions to the starting values for the full model.

By default, PROC GLIMMIX obtains starting values by applying null model restrictions to the converged estimates of the full model. Although this is computationally expedient, the method does not always lead to good starting values for the null model, depending on the nature of the model and hypothesis. In particular, when you receive a warning about parameters not specified under falling on the boundary, the RESTART option can be useful.

Values within tolerance of the boundary of the parameter space are considered on the boundary when PROC GLIMMIX examines estimates of nuisance parameters under and determines whether mixture weights and degrees of freedom can be obtained. In certain cases, when parameters not specified under the null hypothesis are on boundaries, the asymptotic distribution of the likelihood ratio statistic is not a mixture of chi-squares (see, for example, case 8 in Self and Liang 1987). The default for is times the machine epsilon; this product is approximately 1E12 on most computers.

produces Wald Z tests for the covariance parameters based on the estimates and asymptotic standard errors in the "Covariance Parameter Estimates" table.

enables you to supply weights for the computation of p-values from chi-square mixtures. See the DF= option for details. Mixture weights are scaled to sum to one.