ScatterplotSmoothers: Smoothers to Draw Lines on Scatterplots
In jonathon-love/car3: Companion to Applied Regression

Description Usage Arguments Details Author(s) See Also Examples

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, log.x, log.y, spread=FALSE, smoother.args, draw=TRUE, offset=0)

loessLine(x, y, col, log.x, log.y, spread=FALSE, smoother.args, draw=TRUE, offset=0)

quantregLine(x, y, col, log.x, log.y, spread=FALSE, smoother.args, draw=TRUE, offset=0)

`x`	$x$ coordinates of points.
`y`	$y$ coordinates of points.
`col`	line color.
`log.x`	`TRUE` if the $x$-axis is logged.
`log.y`	`TRUE` if the $y$-axis is logged.
`spread`	the default is to plot only an estimated mean or median function. If this argument is TRUE, then a measure of spread is also plotted.
`smoother.args`	additional options accepted by the smoother, in the form of a list of named values (see Details below).
`draw`	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.
`offset`	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).

For loessLine the default value is smoother.args=list(lty=1, lwd=2, lty.spread=2, lwd.spread=1, span=2/3 (prior to 11/2016, span was 1/2), degree=2, family="symmetric", iterations=4). The arguments lty and lwd are the type and width respectively of the mean or median smooth, smooth.lty and smooth.lwd 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=1, lwd=2, lty.spread=2, lwd.spread=1, 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=1, lwd=2, lty.spread=2, lwd.spread=1, 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.

John Fox jfox@mcmaster.ca and Sanford Weisberg sandy@umn.edu.

scatterplot, scatterplotMatrix, gam, loess, and rqss.

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))