These smoothers are used to draw nonparametric-regression lines on scatterplots produced by
the scatterplot
, scatterplotMatrix
and several other car functions.
The functions are not meant to
be called directly by the user, although the user can supply options via the smoother.args
argument,
the contents of which vary by the smoother (see Details below). The gamLine
smoother uses the
gam
function in the mgcv package, the loessLine
smoother uses the
loess
function in the stats package, and the quantregLine
smoother uses the
rqss
function in the quantreg package.
gamLine(x, y, col=carPalette()[1], log.x=FALSE, log.y=FALSE, var=FALSE, spread=var,
smoother.args=NULL, draw=TRUE, offset=0)loessLine(x, y, col=carPalette()[1], log.x=FALSE, log.y=FALSE, var=FALSE, spread=var,
smoother.args=NULL, draw=TRUE, offset=0)
quantregLine(x, y, col=carPalette()[1], log.x=FALSE, log.y=FALSE, var=FALSE, spread=var,
smoother.args=NULL, draw=TRUE, offset=0)
horizontal coordinates of points.
vertical coordinates of points.
line color.
should be set to TRUE
(default is FALSE
) if the horizontal axis is logged.
should be set to TRUE
(default is FALSE
) if the vertical axis is logged.
the default is to plot only an estimated mean or median function. If either of these arguments is TRUE, then a measure of variability is also plotted.
additional options accepted by the smoother, in the form of a list of named values (see Details below).
if TRUE, the default, draw the smoother on the currently active graph.
If FALSE, return a list with coordinates x
and y
for the points
that make up the smooth and if requested x.pos, y.pos, x.neg, y.neg
for the
spread smooths.
For use when spread=TRUE
, the vertical axis is sqrt(offset^2 + variance smooth)
.
The function loessLine
is a re-implementation of the loess smoother
that was used in car prior to September 2012. The main enhancement is the ability to
set more arguments through the smoother.args
argument.
The function gamLine
is more general than the loess
fitting
because it allows fitting a generalized additive model using splines. You can specify an error
distribution and link function.
The function quantregLine
fits an additive model using splines with estimation
based on L1 regression for the median and quantile regression if you ask for the spread. It is
likely to be more robust than the other smoothers.
The argument smoother.args
is a list of named elements used to pass
additional arguments to the smoother. As of November, 2016, the smoother is evaluated at an equally spaced grid of 50 points in the range of the horizontal variable. With any of the smoothers you can change to say 100 evaluation points by using the argument smoother.args=list(evaluation=100)
. As of version 3.0-1, the arguments in smoother.args
col.var
, lty.var
and lwd.var
are equivalent to col.spread
, lty.spread
and lwd.spread
, respectively.
For loessLine
the default value is
smoother.args=list(lty.smooth=1, lwd.smooth=2, lty.spread=4, lwd.spread=2, span=2/3 (prior to 11/2016, span was 1/2),
degree=1, family="symmetric", iterations=4)
.
The arguments lty.smooth
and lwd.smooth
are the type and width
respectively of the mean or median smooth, lty.spread
and lwd.spred
are the type and color of the spread smooths if requested.
The arguments span
, degree
and family
are
passed to the loess
function, iterations=4
robustness iterations.
For gamLine
the default is
smoother.args=list(lty.smooth=1, lwd.smooth=2, lty.spread=4, lwd.spread=2,
k=-1, bs="tp", family="gaussian", link=NULL, weights=NULL)
The first four arguments are as for loessLine
. The next two
arguments are passed to the gam
function to control the smoothing:
k=-1
allows gam
to choose the number of splines in the basis
function; bs="tp"
provides the type of spline basis to be used with "tp"
for the default thin-plate splines. The last three arguments allow providing
a family, link and weights as in generalized linear models. See examples
below. The spread argument is ignored unless family="gaussian" and link=NULL.
For quantregLine
the default is
smoother.args=list(lty.smooth=1, lwd.smooth=2, lty.spread=4, lwd.spread=2,
lambda=IQR(x))
. The first four
arguments are as for loessLine
. The last argument is passed to the
qss
function in quantreg
. It is a smoothing
parameter, here a robust estimate of the scale of the horizontal axis variable.
This is an arbitrary choice, and may not work well in all circumstances.
scatterplot
, scatterplotMatrix
, gam
,
loess
, and rqss
.
# NOT RUN {
scatterplot(prestige ~ income, data=Prestige)
scatterplot(prestige ~ income, data=Prestige, smooth=list(smoother=gamLine))
scatterplot(prestige ~ income, data=Prestige,
smooth=list(smoother=quantregLine))
scatterplot(prestige ~ income | type, data=Prestige)
scatterplot(prestige ~ income | type, data=Prestige,
smooth=list(smoother=gamLine))
scatterplot(prestige ~ income | type, data=Prestige,
smooth=list(smoother=quantregLine))
scatterplot(prestige ~ income | type, data=Prestige, smooth=FALSE)
scatterplot(prestige ~ income | type, data=Prestige, spread=TRUE)
scatterplot(prestige ~ income | type, data=Prestige,
smooth=list(smoother=gamLine), spread=TRUE)
scatterplot(prestige ~ income | type, data=Prestige,
smooth=list(smoother=quantregLine), spread=TRUE)
scatterplot(weight ~ repwt | sex, spread=TRUE, data=Davis,
smooth=list(smoother=loessLine))
scatterplot(weight ~ repwt | sex, spread=TRUE, data=Davis,
smooth=list(smoother=gamLine)) # messes up
scatterplot(weight ~ repwt | sex, spread=TRUE, data=Davis,
smooth=list(smoother=quantregLine)) # robust
set.seed(12345)
w <- 1 + rpois(100, 5)
x <- rnorm(100)
p <- 1/(1 + exp(-(x + 0.5*x^2)))
y <- rbinom(100, w, p)
scatterplot(y/w ~ x, smooth=list(smoother=gamLine, family="binomial",
weights=w))
scatterplot(y/w ~ x, smooth=list(smoother=gamLine, family=binomial,
link="probit", weights=w))
scatterplot(y/w ~ x, smooth=list(smoother=loessLine), reg=FALSE)
y <- rbinom(100, 1, p)
scatterplot(y ~ x, smooth=list(smoother=gamLine, family=binomial))
# }
Run the code above in your browser using DataLab