Datafile line qualifiers

Purpose

Common qualifiers

!DEFINE is used to define a variable which is a linear function of factor levels. The variable may then be included in the model.

 !DEFINE Nlin Nitrogen  {-3 -1 1 3}/10

My Basis function

The spline function used in ASReml is the common cubic smoothing spline and has a particular basis function. The !MBF qualifier provides a mechanism for a user to specify the spline covariables based on some other basis function. Read More.

Redefining factors for use in the model

!FAMILY is used to construct a new factor by merging levels of an existing factor, taking the new levels from a file.

 !FAMILY fam(Genotype) Family.txt

!GROUP is used to construct a new factor by merging the levels of an existing factor for subsequent inclusion in the model.

 !GROUP Year  SiteYear  1 2 3 1 2 2 3 4

!SUBSET !SUBGROUP is used to construct a new factor having a subset of the levels of an existing factor, or subset of the variables in a GROUP of variables, for subsequent inclusion in the model.

 !SUBSET Rsite  site  1 3 4 8
 !SUBSET Tr135 Trait 1 3 5 0 0 0

In the second case, padding the list with extra zeros informs ASReml that Trait has six levels: the !SUBSET directive appears before the model line where the (number of) traits are defined. It might appear in a model term us(Tr135).litter so that litter effects are only fitted for traits 1, 3 and 5.

!DEFINE
specifies a particular 'contrast' to be used in fitting a degree of freedom in a fixed factor.

!MBF
sets up a set of covariates read from a file to be used in fitting several degree of freedom in a fixed factor.

The advantage of using these forms (including !DEFINE and !MBF ) over forming new variables/factors using transformations is that in prediction, ASReml knows the relationship between the levels/values of the derived variable to the original variable. It does not retain that connection when variables are derived using transformations.

(De)selecting multivariate mode (!ASUV !ASMV)

!ASUV indicates that while the data structure and linear model are specified in a multivariate form, the residual variance structure is not IDENTITY for units crossed with US for traits. When !ASUV is specified, it is often also necessary to include mv in the linear model and to specify !SIGMAPAR on or before the (leading) model line (original syntax: the !S2==1 qualifier on the R-structure lines).

!ASMV t indicates that while the data structure and linear model are specified in a univariate form, the data is actually arranged in a units by traits form (t nested within n units) and is to be analysed using the residual variance structure IDENTITY for units crossed with US for traits. h2 Missing values in the Design !MVREMOVE
When missing values occur in design variables (as opposed to the response variable), the user must indicate whether to discard the records containing the missing values, or whether to treat the missing values as zeros.

Setting predict points

!PVAL
When fitting polynomials or splines, it is sometimes desirable to predict the response surface at points other than those for which there is data. Use !PVAL to specify these points before the design matrix is formed so that the extra points have the appropriate covariables generated.

!GKRIGE
predicting a grid of points modelled with a correlation structure on a fac(X,Y) model term, use !GKRIGE to indicate how the cordinates are specified.

!FACPOINTS !POLPOINTS

Setting spline knot points

!SPLINE !KNOTS !PPOINTS !SCALE
has a model term spl(x[,k]) which, when fitted as a random term in conjunction with the covariate x fitted as a fixed term generates a cubic smoothing spline model. Several qualifiers modify the defaults with respect to choosing knot points for the spline. See orange tree example.

Defining your own variance structure.

!OWN program
If you want to fit a variance structure not already included in ASReml, write a separate program to create it and its derivatives. The program name is specified with

Other model term qualifiers

!ADJUST !DF !YSS are needed when the data being analysed has been summarized (maybe is means) resulting in a shortage of degrees of freedom or missing residual variance.

Generalized Linear Models

Generalized Linear Mixed Models

!HGLM !HDIST !HLINK Hierarchal Generalized Linear Models

Model transformation

!RBGIBBS Rao-Blackwellized Gibbs

Return to index