completeCgm(data, margins, gauss, design, optData, subset, prior = 1, start = NULL, control = emCgm.control(), contrasts = NULL)
margins
argument is not
provided, then the loglinear part of the model is assumed to be a
saturated model in which all
factor
variables are used to form the
table. If the
gauss
argument is not provided, then all numeric
variables in the data frame are included in the conditional Gaussian
part of the model.
margins
argument,
which provides the names of the variables to use in the discrete part
of the model. If the
gauss
argument is omitted, then all remaining
variables in the matrix are used in the Gaussian part of the model.
list(1:2, 3:4)
would
indicate fitting the 1,2 margin (summing over variables 3 and 4) and
the 3,4 margin in a four-way table. This same model can be specified
using the names of the variables (e.g., list(c("V1", "V2"), c("V3",
"V4"))), or using formula notation, as in
margins = ~V1:V2 + V3:V4
.
margins
is not specified, a saturated model is fitted.
data
, argument
margins
must be
specified. When a data frame is input and argument
margins
is
missing, then a saturated model involving all factor variables is fitted.
"missmodel"
object is input, then if
margins
is not
given, argument
margins
defaults to the margins specified in the
call
statement of the input
"missmodel"
object.
c(1, 2, 4)
; as a
vector of variable names, e.g.
c("V1", "V2", "V4")
; and using
formula notation, e.g.
~V1+V2+V4
. If argument
gauss
is
omitted, then all numeric variables (which do not appear in argument
margins
) are used in the multivariate gaussian model.
optData
. Optionally, an
ncell
by
m
matrix may be input directly as the design matrix.
i=1, ..., ncell
denote the cells in the loglinear model, and let
mu(i)
denote the vector of numeric variable means in cell
i
. Then
the formula
design
provides the design matrix for predicting the
cell means. As an example, let
"V1"
and
"V2"
be the names of the
factor variables, and let
"age"
be a vector giving an average age
for the subjects in each cell. Then formula
design=~V1+V2
indicates
a main effect model for the cell means, while
design=~V1 + V2 + age
indicates a main effect model for the cell means, adjusted for average
cell age.
ncell
by
m
matrix may be input. In this case, the
regression model is obtained as a linear function of the columns of the
input matrix.
design
is not specified, then the design matrix is taken to be an
identity matrix.
ncell
rows containing predictors to be used in
computing the
design
matrix. In the example given in the description
for argument
design
, the variable
age
would be input in argument
optData
.
data
is a data frame,
this expression may use variables in the data frame.
"priorLoglin"
,
or a vector of hyperparameters.
"ml"
(maximum likelihood),
"noninformative"
, and
"data.dependent"
. String matching is used,
so the characters
"m"
,
"n"
, or
"d"
are sufficient. The values
of the hyperparameters change with the algorithm (see
for details). E.g.
"noninformative"
means a common value of 1 for
EM, and a common value of 0.5 for DA.
"priorLoglin"
object is created by routine
priorLoglin
.
dataDepPrior
.
See
for details.
"noninformative"
. When a class
"missmodel"
object is input, any value specified in a previous call has priority
over the default value (but not over any currently specified value).
NA
) when a vector of
hyperparameters is input as argument
prior
.
"missmodel"
object is input and argument
prior
is not
given, then argument
prior
defaults to the prior probabilities
specified in the
call
statement of the input
"missmodel"
object. If these are not specified, then the default (which depends on
the algorithm) is used.
completeCgm
, but is included to conform with other
missing data functions.
completeCgm
, but is included to conform with other
missing data functions.
design
formula. The elements of the list should have the same
name as the variable and should be either a contrast matrix
(specifically, any full-rank matrix with as many rows as there are
levels in the factor), or else a function to compute such a matrix
given the number of levels.
"missmodel"
is returned; see
for details.
The
paramIter
component is of class
cgm
, and is a matrix
whose rows contain parameter estimates. The
algorithm
element contains an object of class
"em"
.
The
completeCgm
function computes Bayes estimates of the
parameters in a multivariate normal model.
Schafer, J. L. (1997), Analysis of Incomplete Multivariate Data, Chapman & Hall, London.
completeCgm(language[,c("LAN", "SEX", "HGPA", "FLAS")], subset = !(is.na(SEX) | is.na(HGPA)), prior = 1) #Equivalent to: mdCgm(language[,c("LAN", "SEX", "HGPA","FLAS")], subset = !(is.na(SEX) | is.na(HGPA)))