Summarize Multiple Response Variables and Make Multipanel Scatter or Dot Plot

Description

Multiple left-hand formula variables along with right-hand side conditioning variables are reshaped into a "tall and thin" data frame if fun is not specified. The resulting raw data can be plotted with the plot method using user-specified panel functions for lattice graphics, typically to make a scatterplot or loess smooths, or both. The Hmisc panel.plsmo function is handy in this context. Instead, if fun is specified, this function takes individual response variables (which may be matrices, as in Surv objects) and creates one or more summary statistics that will be computed while the resulting data frame is being collapsed to one row per condition. The plot method in this case plots a multi-panel dot chart using the lattice dotplot function if panel is not specified to plot. There is an option to print selected statistics as text on the panels. summaryS pays special attention to Hmisc variable annotations: label, units. When panel is specified in addition to fun, a special x-y plot is made that assumes that the x-axis variable (typically time) is discrete. This is used for example to plot multiple quantile intervals as vertical lines next to the main point. A special panel function mvarclPanel is provided for this purpose.

When fun is given and panel is omitted, and the result of fun is a vector of more than one statistic, the first statistic is taken as the main one. Any columns with names not in textonly will figure into the calculation of axis limits. Those in textonly will be printed right under the dot lines in the dot chart. Statistics with names in textplot will figure into limits, be plotted, and printed. pch.stats can be used to specify symbols for statistics after the first column. When fun computed three columns that are plotted, columns two and three are taken as confidence limits for which horizontal "error bars" are drawn. Two levels with different thicknesses are drawn if there are four plotted summary statistics beyond the first.

mbarclPanel is used to draw multiple vertical lines around the main points, such as a series of quantile intervals stratified by x and paneling variables. If mbarclPanel finds a column of an arument yother that is named "se", and if there are exactly two levels to a superpositioning variable, the half-height of the approximate 0.95 confidence interval for the difference between two point estimates is shown, positioned at the midpoint of the two point estimates at an x value. This assume normality of point estimates, and the standard error of the difference is the square root of the sum of squares of the two standard errors. By positioning the intervals in this fashion, a failure of the two point estimates to touch the half-confidence interval is consistent with rejecting the null hypothesis of no difference at the 0.05 level.

medvPanel takes raw data and plots median y vs. x, along with confidence intervals and half-interval for the difference in medians as with mbarclPanel. Quantile intervals are optional. Very transparent vertical violin plots are added by default. Unlike panel.violin, only half of the violin is plotted, and when there are two superpose groups they are side-by-side in different colors.

Usage

summaryS(formula, fun = NULL, data = NULL, subset = NULL,
         na.action = na.retain, continuous=10, ...)

## S3 method for class 'summaryS'
plot(x, formula=NULL, groups=NULL, panel=NULL,
           paneldoesgroups=FALSE, datadensity=NULL, ylab='',
           funlabel=NULL, textonly='n', textplot=NULL,
           digits=3, custom=NULL,
           xlim=NULL, ylim=NULL, cex.strip=1, cex.values=0.5, pch.stats=NULL,
           key=list(columns=length(groupslevels),
             x=.75, y=-.04, cex=.9,
             col=trellis.par.get('superpose.symbol')$col, corner=c(0,1)),
           outerlabels=TRUE, autoarrange=TRUE, scat1d.opts=NULL, ...)

mbarclPanel(x, y, subscripts, groups=NULL, yother, ...)

medvPanel(x, y, subscripts, groups=NULL, violin=TRUE, quantiles=FALSE, ...)

Arguments

Value

a data frame with added attributes for summaryS or a lattice object ready to render for plot

Author(s)

Frank Harrell

See Also

summary, summarize

Examples

# See tests directory file summaryS.r for more examples
n <- 100
set.seed(1)
d <- data.frame(sbp=rnorm(n, 120, 10),
                dbp=rnorm(n, 80, 10),
                age=rnorm(n, 50, 10),
                days=sample(1:n, n, TRUE),
                S1=Surv(2*runif(n)), S2=Surv(runif(n)),
                race=sample(c('Asian', 'Black/AA', 'White'), n, TRUE),
                sex=sample(c('Female', 'Male'), n, TRUE),
                treat=sample(c('A', 'B'), n, TRUE),
                region=sample(c('North America','Europe'), n, TRUE),
                meda=sample(0:1, n, TRUE), medb=sample(0:1, n, TRUE))

d <- upData(d, labels=c(sbp='Systolic BP', dbp='Diastolic BP',
            race='Race', sex='Sex', treat='Treatment',
            days='Time Since Randomization',
            S1='Hospitalization', S2='Re-Operation',
            meda='Medication A', medb='Medication B'),
            units=c(sbp='mmHg', dbp='mmHg', age='Year', days='Days'))

s <- summaryS(age + sbp + dbp ~ days + region + treat,  data=d)
# plot(s)   # 3 pages
plot(s, groups='treat', datadensity=TRUE,
     scat1d.opts=list(lwd=.5, nhistSpike=0))
plot(s, groups='treat', panel=panel.loess, key=list(space='bottom', columns=2),
     datadensity=TRUE, scat1d.opts=list(lwd=.5))

# Make your own plot using data frame created by summaryP
# xyplot(y ~ days | yvar * region, groups=treat, data=s,
#        scales=list(y='free', rot=0))

# Use loess to estimate the probability of two different types of events as
# a function of time
s <- summaryS(meda + medb ~ days + treat + region, data=d)
pan <- function(...)
   panel.plsmo(..., type='l', label.curves=max(which.packet()) == 1,
               datadensity=TRUE)
plot(s, groups='treat', panel=pan, paneldoesgroups=TRUE,
     scat1d.opts=list(lwd=.7), cex.strip=.8)

# Repeat using intervals instead of nonparametric smoother
pan <- function(...)  # really need mobs > 96 to est. proportion
  panel.plsmo(..., type='l', label.curves=max(which.packet()) == 1,
              method='intervals', mobs=5)

plot(s, groups='treat', panel=pan, paneldoesgroups=TRUE, xlim=c(0, 150))


# Demonstrate dot charts of summary statistics
s <- summaryS(age + sbp + dbp ~ region + treat, data=d, fun=mean)
plot(s)
plot(s, groups='treat', funlabel=expression(bar(X)))
# Compute parametric confidence limits for mean, and include sample
# sizes by naming a column "n"

f <- function(x) {
  x <- x[! is.na(x)]
  c(smean.cl.normal(x, na.rm=FALSE), n=length(x))
}
s <- summaryS(age + sbp + dbp ~ region + treat, data=d, fun=f)
plot(s, funlabel=expression(bar(X) %+-% t[0.975] %*% s))
plot(s, groups='treat', cex.values=.65,
     key=list(space='bottom', columns=2,
       text=c('Treatment A:','Treatment B:')))

# For discrete time, plot Harrell-Davis quantiles of y variables across
# time using different line characteristics to distinguish quantiles
d <- upData(d, days=round(days / 30) * 30)
g <- function(y) {
  probs <- c(0.05, 0.125, 0.25, 0.375)
  probs <- sort(c(probs, 1 - probs))
  y <- y[! is.na(y)]
  w <- hdquantile(y, probs)
  m <- hdquantile(y, 0.5, se=TRUE)
  se <- as.numeric(attr(m, 'se'))
  c(Median=as.numeric(m), w, se=se, n=length(y))
}
s <- summaryS(sbp + dbp ~ days + region, fun=g, data=d)
plot(s, panel=mbarclPanel)
plot(s, groups='region', panel=mbarclPanel, paneldoesgroups=TRUE)

# For discrete time, plot median y vs x along with CL for difference,
# using Harrell-Davis median estimator and its s.e., and use violin
# plots

s <- summaryS(sbp + dbp ~ days + region, data=d)
plot(s, groups='region', panel=medvPanel, paneldoesgroups=TRUE)

# Proportions and Wilson confidence limits, plus approx. Gaussian
# based half/width confidence limits for difference in probabilities
g <- function(y) {
  y <- y[!is.na(y)]
  n <- length(y)
  p <- mean(y)
  se <- sqrt(p * (1. - p) / n)
  structure(c(binconf(sum(y), n), se=se, n=n),
            names=c('Proportion', 'Lower', 'Upper', 'se', 'n'))
}
s <- summaryS(meda + medb ~ days + region, fun=g, data=d)
plot(s, groups='region', panel=mbarclPanel, paneldoesgroups=TRUE)