xyplot and dotplot with Matrix Variables to Plot Error Bars and Bands
A utility function Cbind returns the first argument as a vector and
combines all other arguments into a matrix stored as an attribute called
"other". The arguments can be named (e.g.,
Cbind(pressure=y,ylow,yhigh)) or a label attribute may be pre-attached
to the first argument. In either case, the name or label of the first
argument is stored as an attribute "label" of the object returned by
Cbind. Storing other vectors as a matrix attribute facilitates plotting
error bars, etc., as trellis really wants the x- and y-variables to be
vectors, not matrices. If a single argument is given to Cbind and that
argument is a matrix with column dimnames, the first column is taken as the
main vector and remaining columns are taken as "other". A subscript
method for Cbind objects subscripts the other matrix along
with the main y vector.
The xYplot function is a substitute for xyplot that allows for
simulated multi-column y. It uses by default the panel.xYplot and
prepanel.xYplot functions to do the actual work. The method argument
passed to panel.xYplot from xYplot allows you to make error bars, the
upper-only or lower-only portions of error bars, alternating lower-only and
upper-only bars, bands, or filled bands. panel.xYplot decides how to
alternate upper and lower bars according to whether the median y value of
the current main data line is above the median y for all groups of
lines or not. If the median is above the overall median, only the upper
bar is drawn. For bands (but not 'filled bands'), any number of other
columns of y will be drawn as lines having the same thickness, color, and
type as the main data line. If plotting bars, bands, or filled bands and
only one additional column is specified for the response variable, that
column is taken as the half width of a precision interval for y, and the
lower and upper values are computed automatically as y plus or minus the
value of the additional column variable.
When a groups variable is present, panel.xYplot will create a function
in frame 0 (.GlobalEnv in R) called Key that when
invoked will draw a key describing the
groups labels, point symbols, and colors. By default, the key is outside
the graph. For S-Plus, if Key(locator(1)) is specified, the key will appear so that
its upper left corner is at the coordinates of the mouse click. For
R/Lattice the first two arguments of Key (x and y) are fractions
of the page, measured from the lower left corner, and the default
placement is at x=0.05, y=0.95. For R, an optional argument
to sKey, other, may contain a list of arguments to pass to draw.key (see
xyplot for a list of possible arguments, under
the key option).
When method="quantile" is specified, xYplot automatically groups the
x variable into intervals containing a target of nx observations each,
and within each x group computes three quantiles of y and plots these
as three lines. The mean x within each x group is taken as the
x-coordinate. This will make a useful empirical display for large
datasets in which scatterdiagrams are too busy to see patterns of central
tendency and variability. You can also specify a general function of a
data vector that returns a matrix of statistics for the method argument.
Arguments can be passed to that function via a list methodArgs. The
statistic in the first column should be the measure of central tendency.
Examples of useful method functions are those listed under the help file
for summary.formula such as smean.cl.normal.
xYplot can also produce bubble plots. This is done when
size is specified to xYplot. When size is used, a
function sKey is generated for drawing a key to the character
sizes. See the bubble plot example. size can also specify a
vector where the first character of each observation is used as the
plotting symbol, if rangeCex is set to a single cex
value. An optional argument to sKey, other, may contain
a list of arguments to pass to draw.key (see
xyplot for a list of possible arguments, under
the key option). See the bubble plot example.
Dotplot is a substitute for dotplot allowing for a matrix x-variable,
automatic superpositioning when groups is present, and creation of a
Key function. When the x-variable (created by Cbind to simulate a
matrix) contains a total of 3 columns, the first column specifies where the
dot is positioned, and the last 2 columns specify starting and ending
points for intervals. The intervals are shown using line type, width, and
color from the trellis plot.line list. By default, you will usually see a
darker line segment for the low and high values, with the dotted reference
line elsewhere. A good choice of the pch argument for such plots is 3
(plus sign) if you want to emphasize the interval more than the point
estimate. When the x-variable contains a total of 5 columns, the 2nd and
5th columns are treated as the 2nd and 3rd are treated above, and the 3rd
and 4th columns define an inner line segment that will have twice the
thickness of the outer segments. In addition, tick marks separate the outer
and inner segments. This type of display (an example of which appeared in
The Elements of Graphing Data by Cleveland) is very suitable for
displaying two confidence levels (e.g., 0.9 and 0.99) or the 0.05, 0.25,
0.75, 0.95 sample quantiles, for example. For this display, the central
point displays well with a default circle symbol.
setTrellis sets nice defaults for Trellis graphics, assuming that the
graphics device has already been opened if using postscript, etc. By
default, it sets panel strips to blank and reference dot lines to thickness
1 instead of the Trellis default of 2.
numericScale is a utility function that facilitates using
xYplot to
plot variables that are not considered to be numeric but which can readily
be converted to numeric using as.numeric(). numericScale
by default will keep the name of the input variable as a label
attribute for the new numeric variable.
Cbind(...)
xYplot(formula, data = sys.frame(sys.parent()), groups,
subset, xlab=NULL, ylab=NULL, ylim=NULL,
panel=panel.xYplot, prepanel=prepanel.xYplot, scales=NULL,
minor.ticks=NULL, sub=NULL, ...)
panel.xYplot(x, y, subscripts, groups=NULL,
type=if(is.function(method) || method=='quantiles')
'b' else 'p',
method=c("bars", "bands", "upper bars", "lower bars",
"alt bars", "quantiles", "filled bands"),
methodArgs=NULL, label.curves=TRUE, abline,
probs=c(.5,.25,.75), nx=NULL,
cap=0.015, lty.bar=1,
lwd=plot.line$lwd, lty=plot.line$lty, pch=plot.symbol$pch,
cex=plot.symbol$cex, font=plot.symbol$font, col=NULL,
lwd.bands=NULL, lty.bands=NULL, col.bands=NULL,
minor.ticks=NULL, col.fill=NULL,
size=NULL, rangeCex=c(.5,3), ...)
prepanel.xYplot(x, y, ...)
Dotplot(formula, data = sys.frame(sys.parent()), groups, subset,
xlab = NULL, ylab = NULL, ylim = NULL,
panel=panel.Dotplot, prepanel=prepanel.Dotplot,
scales=NULL, xscale=NULL, ...)
prepanel.Dotplot(x, y, ...)
panel.Dotplot(x, y, groups = NULL,
pch = dot.symbol$pch,
col = dot.symbol$col, cex = dot.symbol$cex,
font = dot.symbol$font, abline, ...)
setTrellis(strip.blank=TRUE, lty.dot.line=2, lwd.dot.line=1)
numericScale(x, label=NULL, ...)... |
for Also can be other arguments to pass to |
formula |
a |
x |
|
y |
a vector, or an object created by |
data,subset,ylim,subscripts,groups,type,scales,panel,prepanel,xlab,ylab |
see |
xscale |
allows one to use the default |
method |
defaults to |
methodArgs |
a list containing optional arguments to be passed to the function specified
in |
label.curves |
set to |
abline |
a list of arguments to pass to |
probs |
a vector of three quantiles with the quantile corresponding to the central
line listed first. By default |
nx |
number of target observations for each |
cap |
the half-width of horizontal end pieces for error bars, as a fraction of
the length of the |
lty.bar |
line type for bars |
lwd, lty, pch, cex, font, col |
see |
lty.bands, lwd.bands, col.bands |
used to allow |
minor.ticks |
a list with elements |
sub |
an optional subtitle |
col.fill |
used to override default colors used for the bands in method='filled
bands'. This is a vector when |
size |
a vector the same length as |
rangeCex |
a vector of two values specifying the range in character sizes to use
for the |
strip.blank |
set to |
lty.dot.line |
line type for dot plot reference lines (default = 1 for dotted; use 2 for dotted) |
lwd.dot.line |
line thickness for reference lines for dot plots (default = 1) |
label |
a scalar character string to be used as a variable label after
|
Unlike xyplot, xYplot senses the presence of a groups variable and
automatically invokes panel.superpose instead of panel.xyplot. The same
is true for Dotplot vs. dotplot.
Cbind returns a matrix with attributes. Other functions return standard
trellis results.
plots, and panel.xYplot may create temporary Key and
sKey functions in the session frame.
Frank Harrell
Department of Biostatistics
Vanderbilt University
fh@fharrell.com
Madeline Bauer
Department of Infectious Diseases
University of Southern California School of Medicine
mbauer@usc.edu
# Plot 6 smooth functions. Superpose 3, panel 2.
# Label curves with p=1,2,3 where most separated
d <- expand.grid(x=seq(0,2*pi,length=150), p=1:3, shift=c(0,pi))
xYplot(sin(x+shift)^p ~ x | shift, groups=p, data=d, type='l')
# Use a key instead, use 3 line widths instead of 3 colors
# Put key in most empty portion of each panel
xYplot(sin(x+shift)^p ~ x | shift, groups=p, data=d,
type='l', keys='lines', lwd=1:3, col=1)
# Instead of implicitly using labcurve(), put a
# single key outside of panels at lower left corner
xYplot(sin(x+shift)^p ~ x | shift, groups=p, data=d,
type='l', label.curves=FALSE, lwd=1:3, col=1, lty=1:3)
Key()
# Bubble plots
x <- y <- 1:8
x[2] <- NA
units(x) <- 'cm^2'
z <- 101:108
p <- factor(rep(c('a','b'),4))
g <- c(rep(1,7),2)
data.frame(p, x, y, z, g)
xYplot(y ~ x | p, groups=g, size=z)
Key(other=list(title='g', cex.title=1.2)) # draw key for colors
sKey(.2,.85,other=list(title='Z Values', cex.title=1.2))
# draw key for character sizes
# Show the median and quartiles of height given age, stratified
# by sex and race. Draws 2 sets (male, female) of 3 lines per panel.
# xYplot(height ~ age | race, groups=sex, method='quantiles')
# Examples of plotting raw data
dfr <- expand.grid(month=1:12, continent=c('Europe','USA'),
sex=c('female','male'))
set.seed(1)
dfr <- upData(dfr,
y=month/10 + 1*(sex=='female') + 2*(continent=='Europe') +
runif(48,-.15,.15),
lower=y - runif(48,.05,.15),
upper=y + runif(48,.05,.15))
xYplot(Cbind(y,lower,upper) ~ month,subset=sex=='male' & continent=='USA',
data=dfr)
xYplot(Cbind(y,lower,upper) ~ month|continent, subset=sex=='male',data=dfr)
xYplot(Cbind(y,lower,upper) ~ month|continent, groups=sex, data=dfr); Key()
# add ,label.curves=FALSE to suppress use of labcurve to label curves where
# farthest apart
xYplot(Cbind(y,lower,upper) ~ month,groups=sex,
subset=continent=='Europe', data=dfr)
xYplot(Cbind(y,lower,upper) ~ month,groups=sex, type='b',
subset=continent=='Europe', keys='lines',
data=dfr)
# keys='lines' causes labcurve to draw a legend where the panel is most empty
xYplot(Cbind(y,lower,upper) ~ month,groups=sex, type='b', data=dfr,
subset=continent=='Europe',method='bands')
xYplot(Cbind(y,lower,upper) ~ month,groups=sex, type='b', data=dfr,
subset=continent=='Europe',method='upper')
label(dfr$y) <- 'Quality of Life Score'
# label is in Hmisc library = attr(y,'label') <- 'Quality\dots'; will be
# y-axis label
# can also specify Cbind('Quality of Life Score'=y,lower,upper)
xYplot(Cbind(y,lower,upper) ~ month, groups=sex,
subset=continent=='Europe', method='alt bars',
offset=unit(.1,'inches'), type='b', data=dfr)
# offset passed to labcurve to label .4 y units away from curve
# for R (using grid/lattice), offset is specified using the grid
# unit function, e.g., offset=unit(.4,'native') or
# offset=unit(.1,'inches') or unit(.05,'npc')
# The following example uses the summarize function in Hmisc to
# compute the median and outer quartiles. The outer quartiles are
# displayed using "error bars"
set.seed(111)
dfr <- expand.grid(month=1:12, year=c(1997,1998), reps=1:100)
month <- dfr$month; year <- dfr$year
y <- abs(month-6.5) + 2*runif(length(month)) + year-1997
s <- summarize(y, llist(month,year), smedian.hilow, conf.int=.5)
xYplot(Cbind(y,Lower,Upper) ~ month, groups=year, data=s,
keys='lines', method='alt', type='b')
# Can also do:
s <- summarize(y, llist(month,year), quantile, probs=c(.5,.25,.75),
stat.name=c('y','Q1','Q3'))
xYplot(Cbind(y, Q1, Q3) ~ month, groups=year, data=s,
type='b', keys='lines')
# Or:
xYplot(y ~ month, groups=year, keys='lines', nx=FALSE, method='quantile',
type='b')
# nx=FALSE means to treat month as a discrete variable
# To display means and bootstrapped nonparametric confidence intervals
# use:
s <- summarize(y, llist(month,year), smean.cl.boot)
s
xYplot(Cbind(y, Lower, Upper) ~ month | year, data=s, type='b')
# Can also use Y <- cbind(y, Lower, Upper); xYplot(Cbind(Y) ~ ...)
# Or:
xYplot(y ~ month | year, nx=FALSE, method=smean.cl.boot, type='b')
# This example uses the summarize function in Hmisc to
# compute the median and outer quartiles. The outer quartiles are
# displayed using "filled bands"
s <- summarize(y, llist(month,year), smedian.hilow, conf.int=.5)
# filled bands: default fill = pastel colors matching solid colors
# in superpose.line (this works differently in R)
xYplot ( Cbind ( y, Lower, Upper ) ~ month, groups=year,
method="filled bands" , data=s, type="l")
# note colors based on levels of selected subgroups, not first two colors
xYplot ( Cbind ( y, Lower, Upper ) ~ month, groups=year,
method="filled bands" , data=s, type="l",
subset=(year == 1998 | year == 2000), label.curves=FALSE )
# filled bands using black lines with selected solid colors for fill
xYplot ( Cbind ( y, Lower, Upper ) ~ month, groups=year,
method="filled bands" , data=s, label.curves=FALSE,
type="l", col=1, col.fill = 2:3)
Key(.5,.8,col = 2:3) #use fill colors in key
# A good way to check for stable variance of residuals from ols
# xYplot(resid(fit) ~ fitted(fit), method=smean.sdl)
# smean.sdl is defined with summary.formula in Hmisc
# Plot y vs. a special variable x
# xYplot(y ~ numericScale(x, label='Label for X') | country)
# For this example could omit label= and specify
# y ~ numericScale(x) | country, xlab='Label for X'
# Here is an example of using xYplot with several options
# to change various Trellis parameters,
# xYplot(y ~ x | z, groups=v, pch=c('1','2','3'),
# layout=c(3,1), # 3 panels side by side
# ylab='Y Label', xlab='X Label',
# main=list('Main Title', cex=1.5),
# par.strip.text=list(cex=1.2),
# strip=function(\dots) strip.default(\dots, style=1),
# scales=list(alternating=FALSE))
#
# Dotplot examples
#
s <- summarize(y, llist(month,year), smedian.hilow, conf.int=.5)
setTrellis() # blank conditioning panel backgrounds
Dotplot(month ~ Cbind(y, Lower, Upper) | year, data=s)
# or Cbind(\dots), groups=year, data=s
# Display a 5-number (5-quantile) summary (2 intervals, dot=median)
# Note that summarize produces a matrix for y, and Cbind(y) trusts the
# first column to be the point estimate (here the median)
s <- summarize(y, llist(month,year), quantile,
probs=c(.5,.05,.25,.75,.95), type='matrix')
Dotplot(month ~ Cbind(y) | year, data=s)
# Use factor(year) to make actual years appear in conditioning title strips
# Plot proportions and their Wilson confidence limits
set.seed(3)
d <- expand.grid(continent=c('USA','Europe'), year=1999:2001,
reps=1:100)
# Generate binary events from a population probability of 0.2
# of the event, same for all years and continents
d$y <- ifelse(runif(6*100) <= .2, 1, 0)
s <- with(d,
summarize(y, llist(continent,year),
function(y) {
n <- sum(!is.na(y))
s <- sum(y, na.rm=TRUE)
binconf(s, n)
}, type='matrix')
)
Dotplot(year ~ Cbind(y) | continent, data=s, ylab='Year',
xlab='Probability')
# Dotplot(z ~ x | g1*g2)
# 2-way conditioning
# Dotplot(z ~ x | g1, groups=g2); Key()
# Key defines symbols for g2
# If the data are organized so that the mean, lower, and upper
# confidence limits are in separate records, the Hmisc reShape
# function is useful for assembling these 3 values as 3 variables
# a single observation, e.g., assuming type has values such as
# c('Mean','Lower','Upper'):
# a <- reShape(y, id=month, colvar=type)
# This will make a matrix with 3 columns named Mean Lower Upper
# and with 1/3 as many rows as the original dataPlease choose more modern alternatives, such as Google Chrome or Mozilla Firefox.