RANDOM Statement 
The matrix is constructed exactly like the matrix for the fixed effects, and the matrix is constructed to correspond to the effects constituting . The structures of and are defined by using the TYPE= option described on . The random effects can be classification or continuous effects, and multiple RANDOM statements are possible.
Some reserved keywords have special significance in the randomeffects list. You can specify INTERCEPT (or INT) as a random effect to indicate the intercept. PROC GLIMMIX does not include the intercept in the RANDOM statement by default as it does in the MODEL statement. You can specify the _RESIDUAL_ keyword (or RESID, RESIDUAL, _RESID_) before the option slash (/) to indicate a residualtype (Rside) random component that defines the matrix. Basically, the _RESIDUAL_ keyword takes the place of the randomeffect if you want to specify Rside variances and covariance structures. These keywords take precedence over variables in the data set with the same name. If your data or the covariance structure requires that an effect is specified, you can use the RESIDUAL option to instruct the GLIMMIX procedure to model the Rside variances and covariances.
In order to add an overdispersion component to the variance function, simply specify a single residual random component. For example, the following statements fit a polynomial Poisson regression model with overdispersion. The variance function is replaced by :
proc glimmix; model count = x x*x / dist=poisson; random _residual_; run;
Table 40.12 summarizes important options in the RANDOM statement. All options are subsequently discussed in alphabetical order.
Option 
Description 

Construction of Covariance Structure 

determines coordinate association for Gside spatial structures with repeat levels 

varies covariance parameters by groups 

specifies a data set with coefficient matrices for TYPE= LIN 

eliminates columns in corresponding to missing values 

designates a covariance structure as Rside 

identifies the subjects in the model 

specifies the covariance structure 

Mixed Model Smoothing 

displays spline knots 

specifies the upper limit for knot construction 

specifies the method for constructing knots for radial smoother and penalized Bsplines 

specifies the lower limit for knot construction 

Statistical Output 

determines the confidence level () 

requests confidence limits for predictors of random effects 

displays the estimated matrix 

displays the Cholesky root (lower) of the estimated matrix 

displays the inverse Cholesky root (lower) of the estimated matrix 

displays the correlation matrix that corresponds to the estimated matrix 

displays the inverse of the estimated matrix 

displays solutions of the Gside random effects 

displays blocks of the estimated matrix 

displays the lowertriangular Cholesky root of blocks of the estimated matrix 

displays the inverse Cholesky root of blocks of the estimated matrix 

displays the correlation matrix corresponding to blocks of the estimated matrix 

displays the inverse of the blocks of the estimated matrix 
You can specify the following options in the RANDOM statement after a slash (/).
requests that a ttype confidence interval with confidence level be constructed for the predictors of Gside random effects in this statement. The value of number must be between 0 and 1; the default is 0.05. Specifying the ALPHA= option implies the CL option.
requests that ttype confidence limits be constructed for each of the predictors of Gside random effects in this statement. The confidence level is 0.95 by default; this can be changed with the ALPHA= option. The CL option implies the SOLUTION option.
requests that the estimated matrix be displayed for Gside random effects associated with this RANDOM statement. PROC GLIMMIX displays blanks for values that are 0.
displays the lowertriangular Cholesky root of the estimated matrix for Gside random effects.
displays the inverse Cholesky root of the estimated matrix for Gside random effects.
determines how the GLIMMIX procedure associates coordinates for TYPE=SP() covariance structures with effect levels for Gside random effects. In these covariance structures, you specify one or more variables that identify the coordinates of a data point. The levels of classification variables, on the other hand, can occur multiple times for a particular subject. For example, in the following statements the same level of A can occur multiple times, and the associated values of x might be different:
proc glimmix; class A B; model y = B; random A / type=sp(pow)(x); run;
The GCOORD=LAST option determines the coordinates for a level of the random effect from the last observation associated with the level. Similarly, the GCOORD=FIRST and GCOORD=MEAN options determine the coordinate from the first observation and from the average of the observations. Observations not used in the analysis are not considered in determining the first, last, or average coordinate. The default is GCOORD=LAST.
displays the correlation matrix that corresponds to the estimated matrix for Gside random effects.
displays the inverse of the estimated matrix for Gside random effects.
identifies groups by which to vary the covariance parameters. Each new level of the grouping effect produces a new set of covariance parameters. Continuous variables and computed variables are permitted as group effects. PROC GLIMMIX does not sort by the values of the continuous variable; rather, it considers the data to be from a new group whenever the value of the continuous variable changes from the previous observation. Using a continuous variable decreases execution time for models with a large number of groups and also prevents the production of a large "Class Levels Information" table.
Specifying a GROUP effect can greatly increase the number of estimated covariance parameters, which can adversely affect the optimization process.
displays the number and coordinates of the knots as determined by the KNOTMETHOD= option.
provides upper limits for the values of random effects used in the construction of knots for TYPE=RSMOOTH. The items in numberlist correspond to the random effects of the radial smooth. If the KNOTMAX= option is not specified, or if the value associated with a particular random effect is set to missing, the maximum is based on the values in the data set for KNOTMETHOD=EQUAL or KNOTMETHOD=KDTREE, and is based on the values in the knot data set for KNOTMETHOD=DATA.
determines the method of constructing knots for the radial smoother fit with the TYPE=RSMOOTH covariance structure and the TYPE=PSPLINE covariance structure.
Unless you select the TYPE=RSMOOTH or TYPE=PSPLINE covariance structure, the KNOTMETHOD= option has no effect. The default for TYPE=RSMOOTH is KNOTMETHOD=KDTREE. For TYPE=PSPLINE, only equally spaced knots are used and you can use the optional numberlist argument of KNOTMETHOD=EQUAL to determine the number of interior knots for TYPE=PSPLINE.
Knot Construction for TYPE=RSMOOTH
PROC GLIMMIX fits a lowrank smoother, meaning that the number of knots is considerably less than the number of observations. By default, PROC GLIMMIX determines the knot locations based on the vertices of a kd tree (Friedman, Bentley, and Finkel 1977; Cleveland and Grosse 1991). The kd tree is a tree data structure that is useful for efficiently determining the nearest neighbors of a point. The kd tree also can be used to obtain a grid of points that adapts to the configuration of the data. The process starts with a hypercube that encloses the values of the random effects. The space is then partitioned recursively by splitting cells at the median of the data in the cell for the random effect. The procedure is repeated for all cells that contain more than a specified number of points, . The value is called the bucket size.
The kd tree is thus a division of the data into cells such that cells representing leaf nodes contain at most values. You control the building of the kd tree through the BUCKET= treeoption. You control the construction of knots from the cell coordinates of the tree with the other options as follows.
determines the bucket size . A larger bucket size will result in fewer knots. For kd trees in more than one dimension, the correspondence between bucket size and number of knots is difficult to determine. It depends on the data configuration and on other suboptions. In the multivariate case, you might need to try out different bucket sizes to obtain the desired number of knots. The default value of number is 4 for univariate trees (a single random effect) and in the multidimensional case.
specifies whether the knots are based on vertices of the tree cells or the centroid. The two possible values of type are VERTEX and CENTER. The default is KNOTTYPE=VERTEX. For multidimensional smoothing, such as smoothing across irregularly shaped spatial domains, the KNOTTYPE=CENTER option is useful to move knot locations away from the bounding hypercube toward the convex hull.
specifies that knot coordinates are the coordinates of the nearest neighbor of either the centroid or vertex of the cell, as determined by the KNOTTYPE= suboption.
displays details about the construction of the kd tree, such as the cell splits and the split values.
See the section Knot Selection for a detailed example of how the specification of the bucket size translates into the construction of a kd tree and the spline knots.
The KNOTMETHOD=EQUAL option enables you to define a regular grid of knots. By default, PROC GLIMMIX constructs 10 knots for onedimensional smooths and 5 knots in each dimension for smoothing in higher dimensions. You can specify a different number of knots with the optional numberlist. Missing values in the numberlist are replaced with the default values. A minimum of two knots in each dimension is required. For example, the following statements use a rectangular grid of 35 knots, five knots for x1 combined with seven knots for x2:
proc glimmix; model y=; random x1 x2 / type=rsmooth knotmethod=equal(5 7); run;
When you use the NOFIT option in the PROC GLIMMIX statement, the GLIMMIX procedure computes the knots but does not fit the model. This can be useful if you want to compare knot selections with different suboptions of KNOTMETHOD=KDTREE. Suppose you want to determine the number of knots based on a particular bucket size. The following statements compute and display the knots in a bivariate smooth, constructed from nearest neighbors of the vertices of a kd tree with bucket size 10:
proc glimmix nofit; model y = Latitude Longitude; random Latitude Longitude / type=rsmooth knotmethod=kdtree(knottype=vertex nearest bucket=10) knotinfo; run;
You can specify a data set that contains variables whose values give the knot coordinates with the KNOTMETHOD=DATA option. The data set must contain numeric variables with the same name as the radial smoothing randomeffects. PROC GLIMMIX uses only the unique knot coordinates in the knot data set. This option is useful to provide knot coordinates different from those that can be produced from a kd tree. For example, in spatial problems where the domain is irregularly shaped, you might want to determine knots by a spacefilling algorithm. The following SAS statements invoke the OPTEX procedure to compute 45 knots that uniformly cover the convex hull of the data locations (see Chapter 30, "Introduction to the OPTEX Procedure," (SAS/QC User’s Guide) and Chapter 31, "Details of the OPTEX Procedure," (SAS/QC User’s Guide) for details about the OPTEX procedure).
proc optex coding=none; model latitude longitude / noint; generate n=45 criterion=u method=m_fedorov; output out=knotdata; run; proc glimmix; model y = Latitude Longitude; random Latitude Longitude / type=rsmooth knotmethod=data(knotdata); run;
Knot Construction for TYPE=PSPLINE
Only evenly spaced knots are supported when you fit penalized Bsplines with the GLIMMIX procedure. For the TYPE=PSPLINE covariance structure, the numberlist argument specifies the number of interior knots, the default is . Suppose that and denote the smallest and largest values, respectively. For a Bspline of degree (de Boor 2001), the interior knots are supplemented with exterior knots below and exterior knots above . PROC GLIMMIX computes the location of these knots as follows. Let , then interior knots are placed at
The exterior knots are also evenly spaced with step size and start at times the machine epsilon. At least one interior knot is required.
provides lower limits for the values of random effects used in the construction of knots for TYPE=RSMOOTH. The items in numberlist correspond to the random effects of the radial smooth. If the KNOTMIN= option is not specified, or if the value associated with a particular random effect is set to missing, the minimum is based on the values in the data set for KNOTMETHOD=EQUAL or KNOTMETHOD=KDTREE, and is based on the values in the knot data set for KNOTMETHOD=DATA.
reads the coefficient matrices for the TYPE=LIN(q) option. You can specify the LDATA= data set in a sparse or dense form. In the sparse form the data set must contain the numeric variables Parm, Row, Col, and Value. The Parm variable contains the indices of the matrices. The Row and Col variables identify the position within a matrix and the Value variable contains the matrix element. Values not specified for a particular row and column are set to zero. Missing values are allowed in the Value column of the LDATA= data set; these values are also replaced by zeros. The sparse form is particularly useful if the matrices have only a few nonzero elements.
In the dense form the LDATA= data set contains the numeric variables Parm and Row (with the same function as above), in addition to the numeric variables Col1–Colq. If you omit one or more of the Col1–Colq variables from the data set, zeros are assumed for the respective rows and columns of the matrix. Missing values for Col1–Colq are ignored in the dense form.
The GLIMMIX procedure assumes that the matrices are symmetric. In the sparse LDATA= form you do not need to specify offdiagonal elements in position and . One of them is sufficient. Rowcolumn indices are converted in both storage forms into positions in lower triangular storage. If you specify multiple values in row and column of a particular matrix, only the last value is used. For example, assume you are specifying elements of a matrix. The lower triangular storage of matrix defined by
data ldata; input parm row col value; datalines; 3 2 1 2 3 1 2 5 ;
is
eliminates the columns in corresponding to missing levels of random effects involving CLASS variables. By default, these columns are included in . It is sufficient to specify the NOFULLZ option on any Gside RANDOM statement.
specifies that the random effects listed in this statement be Rside effects. You use the RESIDUAL option in the RANDOM statement if the nature of the covariance structure requires you to specify an effect. For example, if it is necessary to order the columns of the Rside AR(1) covariance structure by the time variable, you can use the RESIDUAL option as in the following statements:
class time id; random time / subject=id type=ar(1) residual;
requests that the solution for the randomeffects parameters be produced, if the statement defines Gside random effects.
The numbers displayed in the Std Err Pred column of the "Solution for Random Effects" table are not the standard errors of the displayed in the Estimate column; rather, they are the square roots of the prediction errors , where is the predictor of the th random effect and is the ith random effect. In pseudolikelihood methods that are based on linearization, these EBLUPs are the estimated best linear unbiased predictors in the linear mixed pseudomodel. In models fit by maximum likelihood by using the Laplace approximation or by using adaptive quadrature, the SOLUTION option displays the empirical Bayes estimates (EBE) of .
identifies the subjects in your generalized linear mixed model. Complete independence is assumed across subjects. Specifying a subject effect is equivalent to nesting all other effects in the RANDOM statement within the subject effect.
Continuous variables and computed variables are permitted with the SUBJECT= option. PROC GLIMMIX does not sort by the values of the continuous variable but considers the data to be from a new subject whenever the value of the continuous variable changes from the previous observation. Using a continuous variable can decrease execution time for models with a large number of subjects and also prevents the production of a large "Class Levels Information" table.
specifies the covariance structure of for Gside effects and the covariance structure of for Rside effects.
Although a variety of structures are available, many applications call for either simple diagonal (TYPE=VC) or unstructured covariance matrices. The TYPE=VC (variance components) option is the default structure, and it models a different variance component for each random effect. It is recommended to model unstructured covariance matrices in terms of their Cholesky parameterization (TYPE=CHOL) rather than TYPE=UN.
If you want different covariance structures in different parts of , you must use multiple RANDOM statements with different TYPE= options.
Valid values for covariancestructure are as follows. Examples are shown in Table 40.14.
The variances and covariances in the formulas that follow in the TYPE= descriptions are expressed in terms of generic random variables and . They represent the Gside random effects or the residual random variables for which the or matrices are constructed.
specifies a firstorder antedependence structure (Kenward 1987; Patel 1991) parameterized in terms of variances and correlation parameters. If ordered random variables have a firstorder antedependence structure, then each , , is independent of all other , given . This Markovian structure is characterized by its inverse variance matrix, which is tridiagonal. Parameterizing an ANTE(1) structure for a random vector of size requires parameters: variances and correlation parameters . The covariances among random variables and are then constructed as
PROC GLIMMIX constrains the correlation parameters to satisfy . For variableorder antedependence models see Macchiavelli and Arnold (1994).
specifies a firstorder autoregressive structure,
The values and are derived for the th and th observations, respectively, and are not necessarily the observation numbers. For example, in the following statements the values correspond to the class levels for the time effect of the th and th observation within a particular subject:
proc glimmix; class time patient; model y = x x*x; random time / sub=patient type=ar(1); run;
PROC GLIMMIX imposes the constraint for stationarity.
specifies a heterogeneous firstorder autoregressive structure,
with . This covariance structure has the same correlation pattern as the TYPE=AR(1) structure, but the variances are allowed to differ.
specifies the firstorder autoregressive movingaverage structure,
Here, is the autoregressive parameter, models a movingaverage component, and is a scale parameter. In the notation of Fuller (1976, p. 68), and
The example in Table 40.14 and imply that
where and . PROC GLIMMIX imposes the constraints and for stationarity, although for some values of and in this region the resulting covariance matrix is not positive definite. When the estimated value of becomes negative, the computed covariance is multiplied by to account for the negativity.
specifies an unstructured variancecovariance matrix parameterized through its Cholesky root. This parameterization ensures that the resulting variancecovariance matrix is at least positive semidefinite. If all diagonal values are nonzero, it is positive definite. For example, a unstructured covariance matrix can be written as
Without imposing constraints on the three parameters, there is no guarantee that the estimated variance matrix is positive definite. Even if and are nonzero, a large value for can lead to a negative eigenvalue of . The Cholesky root of a positive definite matrix is a lower triangular matrix such that . The Cholesky root of the above matrix can be written as
The elements of the unstructured variance matrix are then simply , , and . Similar operations yield the generalization to covariance matrices of higher orders.
For example, the following statements model the covariance matrix of each subject as an unstructured matrix:
proc glimmix; class sub; model y = x; random _residual_ / subject=sub type=un; run;
The next set of statements accomplishes the same, but the estimated matrix is guaranteed to be nonnegative definite:
proc glimmix; class sub; model y = x; random _residual_ / subject=sub type=chol; run;
The GLIMMIX procedure constrains the diagonal elements of the Cholesky root to be positive. This guarantees a unique solution when the matrix is positive definite.
The optional order parameter determines how many bands below the diagonal are modeled. Elements in the lower triangular portion of in bands higher than are set to zero. If you consider the resulting covariance matrix , then the order parameter has the effect of zeroing all offdiagonal elements that are at least positions away from the diagonal.
Because of its good computational and statistical properties, the Cholesky root parameterization is generally recommended over a completely unstructured covariance matrix (TYPE=UN). However, it is computationally slightly more involved.
specifies the compoundsymmetry structure, which has constant variance and constant covariance
The compound symmetry structure arises naturally with nested random effects, such as when subsampling error is nested within experimental error. The models constructed with the following two sets of GLIMMIX statements have the same marginal variance matrix, provided is positive:
proc glimmix; class block A; model y = block A; random block*A / type=vc; run;
proc glimmix; class block A; model y = block A; random _residual_ / subject=block*A type=cs; run;
In the first case, the block*A random effect models the Gside experimental error. Because the distribution defaults to the normal, the matrix is of form (see Table 40.15), and is the subsampling error variance. The marginal variance for the data from a particular experimental unit is thus . This matrix is of compound symmetric form.
Hierarchical random assignments or selections, such as subsampling or splitplot designs, give rise to compound symmetric covariance structures. This implies exchangeability of the observations on the subunit, leading to constant correlations between the observations. Compound symmetric structures are thus usually not appropriate for processes where correlations decline according to some metric, such as spatial and temporal processes.
Note that Rside compoundsymmetry structures do not impose any constraint on . You can thus use an Rside TYPE=CS structure to emulate a variancecomponent model with unbounded estimate of the variance component.
specifies the heterogeneous compoundsymmetry structure, which is an equicorrelation structure but allows for different variances
specifies the factoranalytic structure with q factors (Jennrich and Schluchter 1986). This structure is of the form , where is a rectangular matrix and is a diagonal matrix with different parameters. When , the elements of in its upperright corner (that is, the elements in the th row and th column for ) are set to zero to fix the rotation of the structure.
specifies a factoranalytic structure with q factors of the form , where is a rectangular matrix and is the dimension of . When , is a lower triangular matrix. When —that is, when the number of factors is less than the dimension of the matrix—this structure is nonnegative definite but not of full rank. In this situation, you can use it to approximate an unstructured covariance matrix.
specifies a covariance structure that satisfies the general HuynhFeldt condition (Huynh and Feldt 1970). For a random vector with elements, this structure has positive parameters and covariances
A covariance matrix generally satisfies the HuynhFeldt condition if it can be written as . The preceding parameterization chooses . Several simpler covariance structures give rise to covariance matrices that also satisfy the HuynhFeldt condition. For example, TYPE=CS, TYPE=VC, and TYPE=UN(1) are nested within TYPE=HF. You can use the COVTEST statement to test the HF structure against one of these simpler structures. Note also that the HF structure is nested within an unstructured covariance matrix.
The TYPE=HF covariance structure can be sensitive to the choice of starting values and the default MIVQUE(0) starting values can be poor for this structure; you can supply your own starting values with the PARMS statement.
specifies a general linear covariance structure with parameters. This structure consists of a linear combination of known matrices that you input with the LDATA= option. Suppose that you want to model the covariance of a random vector of length , and further suppose that are symmetric ) matrices constructed from the information in the LDATA= data set. Then,
where denotes the element in row , column of matrix .
Linear structures are very flexible and general. You need to exercise caution to ensure that the variance matrix is positive definite. Note that PROC GLIMMIX does not impose boundary constraints on the parameters of a general linear covariance structure. For example, if classification variable A has 6 levels, the following statements fit a variance component structure for the random effect without boundary constraints:
data ldata; retain parm 1 value 1; do row=1 to 6; col=row; output; end; run; proc glimmix data=MyData; class A B; model Y = B; random A / type=lin(1) ldata=ldata; run;
requests that PROC GLIMMIX form a Bspline basis and fits a penalized Bspline (Pspline, Eilers and Marx 1996) with random spline coefficients. This covariance structure is available only for Gside random effects and only a single continuous random effect can be specified with TYPE=PSPLINE. As for TYPE=RSMOOTH, PROC GLIMMIX forms a modified matrix and fits a mixed model in which the random variables associated with the columns of are independent with a common variance. The matrix is constructed as follows.
Denote as the matrix of Bsplines of degree and denote as the matrix of thorder differences. For example, for ,
Then, the matrix used in fitting the mixed model is the matrix
The construction of the Bspline knots is controlled with the KNOTMETHOD= EQUAL() option and the DEGREE= suboption of TYPE=PSPLINE. The total number of knots equals the number of equally spaced interior knots plus knots at the low end and knots at the high end. The number of columns in the Bspline basis equals . By default, the interior knots exclude the minimum and maximum of the randomeffect values and are based on equally spaced intervals. Suppose and are the smallest and largest randomeffect values; then interior knots are placed at
In addition, evenly spaced exterior knots are placed below and exterior knots are placed above . The exterior knots are evenly spaced and start at times the machine epsilon. For example, based on the defaults , , the following statements lead to 26 total knots and 21 columns in , , , :
proc glimmix; model y = x; random x / type=pspline knotmethod=equal(20); run;
Details about the computation and properties of Bsplines can be found in de Boor (2001).
You can extend or limit the range of the knots with the KNOTMIN= and KNOTMAX= options. Table 40.13 lists some of the parameters that control this covariance type and their relationships.
Parameter 
Description 

Degree of Bspline, default 

Order of differencing in construction of , default 

Number of interior knots, default 

Total number of knots 

Number of columns in Bspline basis 

Number of columns in 
You can specify the following options for TYPE=PSPLINE:
specifies the degree of the Bspline. The default is .
specifies the order of the differencing matrix . The default and maximum is .
specifies a radial smoother covariance structure for Gside random effects. This results in an approximate lowrank thinplate spline where the smoothing parameter is obtained by the estimation method selected with the METHOD= option of the PROC GLIMMIX statement. The smoother is based on the automatic smoother in Ruppert, Wand, and Carroll (2003, Chapter 13.4–13.5), but with a different method of selecting the spline knots. See the section Radial Smoothing Based on Mixed Models for further details about the construction of the smoother and the knot selection.
Radial smoothing is possible in one or more dimensions. A univariate smoother is obtained with a single random effect, while multiple random effects in a RANDOM statement yield a multivariate smoother. Only continuous random effects are permitted with this covariance structure. If denotes the number of continuous random effects in the RANDOM statement, then the covariance structure of the random effects is determined as follows. Suppose that denotes the vector of random effects for the th observation. Let denote the vector of knot coordinates, , and is the total number of knots. The Euclidean distance between the knots is computed as
and the distance between knots and effects is computed as
The matrix for the GLMM is constructed as
where the matrix has typical element
and the matrix has typical element
The exponent in these expressions equals , where the optional value corresponds to the derivative penalized in the thinplate spline. A larger value of will yield a smoother fit. The GLIMMIX procedure requires and chooses by default if and otherwise. The NOLOG option removes the and terms from the computation of the and matrices when is even; this yields invariance under rescaling of the coordinates.
Finally, the components of are assumed to have equal variance . The "smoothing parameter" of the lowrank spline is related to the variance components in the model, . See Ruppert, Wand, and Carroll (2003) for details. If the conditional distribution does not provide a scale parameter , you can add a single Rside residual parameter.
The knot selection is controlled with the KNOTMETHOD= option. The GLIMMIX procedure selects knots automatically based on the vertices of a kd tree or reads knots from a data set that you supply. See the section Radial Smoothing Based on Mixed Models for further details on radial smoothing in the GLIMMIX procedure and its connection to a mixed model formulation.
models an exponential spatial or temporal covariance structure, where the covariance between two observations depends on a distance metric . The clist contains the names of the numeric variables used as coordinates to determine distance. For a stochastic process in , there are elements in clist. If the vectors of coordinates for observations and are and , then PROC GLIMMIX computes the Euclidean distance
The covariance between two observations is then
The parameter is not what is commonly referred to as the range parameter in geostatistical applications. The practical range of a (secondorder stationary) spatial process is the distance at which the correlations fall below 0.05. For the SP(EXP) structure, this distance is . PROC GLIMMIX constrains to be positive.
models a gaussian covariance structure,
See TYPE=SP(EXP) for the computation of the distance . The parameter is related to the range of the process as follows. If the practical range is defined as the distance at which the correlations fall below 0.05, then . PROC GLIMMIX constrains to be positive. See TYPE=SP(EXP) for the computation of the distance from the variables specified in clist.
models a covariance structure in the Matérn class of covariance functions (Matérn 1986). The covariance is expressed in the parameterization of Handcock and Stein (1993) and Handcock and Wallis (1994); it can be written as
The function is the modified Bessel function of the second kind of (real) order . The smoothness (continuity) of a stochastic process with covariance function in the Matérn class increases with . This class thus enables datadriven estimation of the smoothness properties of the process. The covariance is identical to the exponential model for (TYPE=SP(EXP)(clist)), while for the model advocated by Whittle (1954) results. As , the model approaches the gaussian covariance structure (TYPE=SP(GAU)(clist)).
Note that the MIXED procedure offers covariance structures in the Matérn class in two parameterizations, TYPE=SP(MATERN) and TYPE=SP(MATHSW). The TYPE=SP(MAT) in the GLIMMIX procedure is equivalent to TYPE=SP(MATHSW) in the MIXED procedure.
Computation of the function and its derivatives is numerically demanding; fitting models with Matérn covariance structures can be timeconsuming. Good starting values are essential.
models a power covariance structure,
where . This is a reparameterization of the exponential structure, TYPE=SP(EXP). Specifically, . See TYPE=SP(EXP) for the computation of the distance from the variables specified in clist. When the estimated value of becomes negative, the computed covariance is multiplied by to account for the negativity.
models an anisotropic power covariance structure in dimensions, provided that the coordinate list clist has elements. If denotes the coordinate for the th observation of the th variable in clist, the covariance between two observations is given by
Note that for , TYPE=SP(POWA) is equivalent to TYPE=SP(POW), which is itself a reparameterization of TYPE=SP(EXP). When the estimated value of becomes negative, the computed covariance is multiplied by to account for the negativity.
models a spherical covariance structure,
The spherical covariance structure has a true range parameter. The covariances between observations are exactly zero when their distance exceeds . See TYPE=SP(EXP) for the computation of the distance from the variables specified in clist.
models a Toeplitz covariance structure. This structure can be viewed as an autoregressive structure with order equal to the dimension of the matrix,
specifies a banded Toeplitz structure,
This can be viewed as a movingaverage structure with order equal to . The specification TYPE=TOEP(1) is the same as , and it can be useful for specifying the same variance component for several effects.
models a Toeplitz covariance structure. The correlations of this structure are banded as the TOEP or TOEP() structures, but the variances are allowed to vary:
The correlation parameters satisfy . If you specify the optional value , the correlation parameters with are set to zero, creating a banded correlation structure. The specification TYPE=TOEPH(1) results in a diagonal covariance matrix with heterogeneous variances.
specifies a completely general (unstructured) covariance matrix parameterized directly in terms of variances and covariances,
The variances are constrained to be nonnegative, and the covariances are unconstrained. This structure is not constrained to be nonnegative definite in order to avoid nonlinear constraints; however, you can use the TYPE=CHOL structure if you want this constraint to be imposed by a Cholesky factorization. If you specify the order parameter , then PROC GLIMMIX estimates only the first bands of the matrix, setting elements in all higher bands equal to 0.
specifies a completely general (unstructured) covariance matrix parameterized in terms of variances and correlations,
where denotes the standard deviation and the correlation is zero when and when , provided the order parameter is given. This structure fits the same model as the TYPE=UN() option, but with a different parameterization. The th variance parameter is . The parameter is the correlation between the th and th measurements; it satisfies . If you specify the order parameter , then PROC GLIMMIX estimates only the first bands of the matrix, setting all higher bands equal to zero.
specifies standard variance components and is the default structure for both Gside and Rside covariance structures. In a Gside covariance structure, a distinct variance component is assigned to each effect. In an Rside structure TYPE=VC is usually used only to add overdispersion effects or with the GROUP= option to specify a heterogeneous variance model.
Description 
Structure 
Example 

Variance 
VC (default) 

Compound 


Heterogeneous 


FirstOrder 


Heterogeneous 


Unstructured 


Banded Main 


Unstructured 


Toeplitz 


Toeplitz with 


Heterogeneous 


Spatial 


FirstOrder 


FirstOrder 


HuynhFeldt 


FirstOrder 

requests that blocks of the estimated marginal variancecovariance matrix be displayed in generalized linear mixed models. This matrix is based on the last linearization as described in the section The Pseudomodel. You can use the valuelist to select the subjects for which the matrix is displayed. If valuelist is not specified, the matrix for the first subject is chosen.
Note that the valuelist refers to subjects as the processing units in the "Dimensions" table. For example, the following statements request that the estimated marginal variance matrix for the second subject be displayed:
proc glimmix; class A B; model y = B; random int / subject=A; random int / subject=A*B v=2; run;
The subject effect for processing in this case is the A effect, because it is contained in the A*B interaction. If there is only a single subject as per the "Dimensions" table, then the V option displays an matrix.
See the section Processing by Subjects for how the GLIMMIX procedure determines the number of subjects in the "Dimensions" table.
The GLIMMIX procedure displays blanks for values that are 0.
displays the lowertriangular Cholesky root of the blocks of the estimated matrix. See the V option for the specification of valuelist.
displays the inverse Cholesky root of the blocks of the estimated matrix. See the V option for the specification of valuelist.
displays the correlation matrix corresponding to the blocks of the estimated matrix. See the V option for the specification of valuelist.
displays the inverse of the blocks of the estimated matrix. See the V option for the specification of valuelist.