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)))