Displays a plot of the data and invites the user to “click” on points judged to lie on the various components.
visualFit(fmla, data=NULL, ncomp, eqVar=FALSE, chsnPts=NULL,
keepPlotVisible = FALSE)A formula specifying the regression model to be fitted.
A list or data frame in which the variables specified by
fmla may be searched for. Variables not found in
data are searched for in the global environment.
Positive integer scalar. The number of components in the mixture which is to be fitted.
Logical scalar; should the error variance be the same for all components? (The alternative is that each component should be allowed to have a different error variance.)
A list with ncomp components each of which is a list
of length two with components x and y. Each
of x and y is a vector of length two, constituting
the \(x\) and \(y\) coordinates respectively of two points
on a line that presumably underlies the corresponding component.
Logical scalar. Should the plot of the data, produced by this
function, be kept visible after the model has been “fitted”?
(Rather than being dismissed by dev.off().)
An object of class "mixreg". (See mixreg().)
Components nsteps and converged are set to NA.
Component data has an extra column groups appended
to it. This column is a factor specifying the components assigned
to the points on the basis of distances from the lines determined
by the chosen points.
The value also has an attribute "chsnPts". This is the
list of points judged by the user to lie on the component lines
(or the value of the chsnPts argument if this was supplied).
If the model involves more than one predictor, or if the specified predictor is a matrix with more than one column, then an error is thrown. This function is intended for use only with one-variable regression.
If there is an intercept in the model, then for each component
(numbered 1 to ncomp) the user is invited to click
on two points judged to lie on a line underlying that
component. If there is no intercept, then the user is invited
to click on a single point for each component, with the origin
(0,0) taken (silently) to be the second point needed to
determine the line.
The fit that this function returns is calculated by assigning a component to each point in the data set, based on which of the visually determined lines that point is closest to.
If eqVar is TRUE then the model is constructed using
a factor, whose entries are these assigned components, as a predictor
(along with the “x” variable in fmla) in a
call to lm(). If code eqVar is FALSE then a
model is fitted separately to each component. (See the code for
details.) “Obviously” the linear coefficient estimates will
be the same in either case. Only the error variance estimates
will differ.
If eq.var is TRUE then the number of parameters in
the model, as used in the calculation of aic, is M =
2*K + (K-1) + 1 = 3*K when there is an intercept term and M
= K + (K-1) + 1 = 2*K when there is no intercept term.
If eq.var is FALSE then the number of parameters
is M = 2*K + (K-1) + K = 4*K - 1 if there is an intercept,
and M = K + (K-1) + K = 3*K - 1 if there is no intercept.
The argument chsnPts allows one to use this function
in a non-interactive session by creating and saving, a
priori, an object to be supplied as the value of chsnPts.
If chsnPts is supplied then the method employed isn't
really “visual”, but presumably the object supplied
would have been created in a visual manner. Be that as it may,
this function is mainly intended to be used visually, that is
without supplying chsnPts.
If chsnPts is not supplied then (“obviously”!!!) this
function can be used only in an interactive session.
mixreg()
# NOT RUN {
vfita <- visualFit(plntsInf ~ aphRel,data=aphids,ncomp=2)
plot(vfita)
vfitk1 <- visualFit(y ~ x, data=kilnAoneOut, ncomp=3)
cp <- attr(vfitk1,"chsnPts")
vfitk2 <- visualFit(y ~ x, data=kilnAoneOut, eqVar=TRUE, chsnPts=cp)
vfitk1$parmat
vfitk2$parmat # Same as above except for the "sigsq" column.
# }
Run the code above in your browser using DataLab