In dcousin3/superb: Summary Plots with Adjusted Error Bars

cat("this will be hidden; use for general initializations.")
library(superb)
library(ggplot2)

By default, superbPlot generates mean plots along with 95% confidence intervals of the mean. However, these choices can be changed.

To change the summary statistics, use the argument statistic =;
To change the interval function, use the argument errorbar =. The abbreviation CI stands for confidence interval; SE stands for standard error.
With CI, to change the confidence level, use the argument gamma =;

The defaults are statistic = "mean", errorbar = "CI", gamma = 0.95. For error bar functions that accept a gamma parameter (e.g., CI, the gamma parameter is automatically transfered to the function). For other functions that do not accept a gamma parameter (e.g., SE), the gamma parameter is unused.

Changing the summary function to be plotted

In what follow, we use GRD() to generate a random dataset with an interaction (see Vignette 6) then make plots varying the statistics displayed.

# shut down 'warnings', 'design' and 'summary' messages
options(superb.feedback = 'none') 

# Generate a random dataset from a (3 x 2) design, entirely within subject.
# The sample size is very small (n=5) and the correlation between scores is high (rho = .8)
dta <- GRD( 
    WSFactors  = "Moment(3): Dose(2)", 
    Effects    = list("Dose*Moment"=custom(0,0,0,1,1,3)), 
    SubjectsPerGroup = 50, 
    Population = list( mean=10, stddev = 5, rho = .80)
)

# a quick function to call superbPlot
makeplot <- function(statfct, errorbarfct, gam, rg, subttl) {
    superb(
        crange(DV.1.1,DV.3.2) ~ ., 
        dta, 
        WSFactors  = c("Moment(3)","Dose(2)"), 
        statistic = statfct, 
        errorbar  = errorbarfct, 
        gamma     = gam, 
        plotStyle = "line",
        adjustments = list(purpose="difference", decorrelation="CM")
    ) + ylab(subttl) + coord_cartesian( ylim = rg )
} 

p1 <- makeplot("mean",      "CI", .95, c(6,14), "Mean +- 95% CI of the mean")
p2 <- makeplot("mean",      "SE", .00, c(6,14), "Mean +- SE of the mean")
p3 <- makeplot("median",    "CI", .95, c(6,14), "Median +- 95% CI of the median")
p4 <- makeplot("fisherskew","CI", .95, c(-2,+2), "Fisher skew +- 95% CI")

library(gridExtra)
p <- grid.arrange(p1,p2,p3,p4, ncol=2)

Determining that a summary function is valid to use in superbPlot

Any summary function can be accepted by superbPlot, as long as it is given within double-quote. The built-in statistics functions such as mean and median can be given. Actually, any descriptive statistics, not just central tendency, can be provided. That includes IQR, mad, etc. To be valid, the function must return a number when given a vector of numbers.

In doubt, you can test if the function is valid for superbPlot with (note the triple colon):

superb:::is.stat.function("mean")

Likewise, any error bar function can be accepted by superbPlot. These functions must be named with two part, separated with a dot "interval function"."descriptive statistic". For example, the function CI.mean is the confidence interval of the mean. Other functions are SE.mean, SE.median, CI.fisherskew, etc. The superb library provides some 20+ such functions. @htc14 and @htc15 reviewed some of these functions.

The error bar functions can be of three types:

a function that returns a width. Standard error functions are example of this type of function. With width function, the error bar extend plus and minus that width around the descriptive statistics.
an interval function. Such functions returns the actual lower and upper limits of the interval and therefore are used as is to draw the bar (i.e., they are not relative to the descriptive statistics). Confidence interval functions are of this type.
"none". This keyword produces an error bar of null width.

The interval function can be tested to see if it exists:

superb:::is.errorbar.function("SE.mean")

To see if a gamma is required for a certain interval function, you can try

superb:::is.gamma.required("SE.mean")

Creating a custom-made descriptive statistic function to be used in superbPlot

As an example, we create from scratch a descriptive statistic function that will be fed to superbPlot. Following @w11 , we implement the 20% trimmed mean. This descriptive statistic is used to estimate the population mean. However, it is said to be a robust statistic as it is less affected by suspicious data. Herein, we use the data from dataFigure1.

    # create a descriptive statistics, the 20% trimmed mean
    trimmedmean    <- function(x) mean(x, trim = 0.2)

    # we can test it with the data from group 1...
    grp1 <- dataFigure1$score[dataFigure1$grp==1]
    grp2 <- dataFigure1$score[dataFigure1$grp==2]
    trimmedmean(grp1)

    # or check that it is a valid statistic function
    superb:::is.stat.function("trimmedmean")

Once we have the function, we can ask for a plot of this function with

superb(
    score ~ grp,
    dataFigure1, 
    statistic = "trimmedmean", errorbar = "none", #HERE the statistic name is given
    plotStyle="line",
    adjustments = list(purpose = "difference"),
    errorbarParams = list(width=0) # so that the null-width error bar is invisible
)+ ylab("20% trimmed mean") +
theme_gray(base_size=10) +
labs(title="20% trimmed mean with \nno error bars") +
coord_cartesian( ylim = c(85,115) )

Creating a custom-made interval function to be used in `superbPlot`

It is also possible to create custom-made confidence interval functions.

Hereafter, we add a confidence interval for the 20% trimmed mean. We use the approach documented in Wilcox [@w11] which requires computing the winsorized standard deviation winsor.sd as available in the psych library [@r20].

library(psych)      # for winsor.sd

CI.trimmedmean <- function(x, gamma = 0.95){
    trim <- 0.2
    g    <- floor(length(x) * 0.4)
    tc   <- qt(1/2+gamma/2, df=(length(x)-g-1) )
    lo   <- tc * winsor.sd(x, trim =0.2) / ((1-2*trim)*sqrt(length(x)))
    c(trimmedmean(x) -lo, trimmedmean(x)+lo)
}

# we test as an example the data from group 1
CI.trimmedmean(grp1)  

# or check that it is a valid interval function
superb:::is.errorbar.function("CI.trimmedmean")

We have all we need to make a plot with error bars!

superb(
    score ~ grp, 
    dataFigure1, 
    statistic = "trimmedmean", errorbar = "CI",
    plotStyle="line",
    adjustments = list(purpose = "difference"),
)+ ylab("20% trimmed mean") +
theme_gray(base_size=10) +
labs(title="20% trimmed mean with \n95% confidence interval of 20% trimmed mean") +
coord_cartesian( ylim = c(85,115) )

The advantage of this measure of central tendancy is that it is a robust estimator. Robust measures are less likely to be adversly affected by suspicious data such as outliers. See @w11 for more on robust estimation.

Creating bootstrap-based confidence intervals

It is also possible to create bootstrap estimates of confidence intervals and integrate these into superb.

The general idea is to subsample with replacement the sample and compute on this subsample the descriptive statistics. This process is repeated a large number of times (here, 10,000) and the quantiles containing, say, 95% of the results is a 95% precision interval (we call it a precision interval as bootstrap estimates the sampling variability, not the predictive variability).

Here, we illustrate this process with the mean. The function must be name "interval function.mean", so we choose to call it myBootstrapCI.mean.

# we define myBootstrapPI which subsample the whole sample, here called X
myBootstrapPI.mean <- function(X, gamma = 0.95) {
  res = c()
  for (i in 1:10000) {
    res[i] <- mean(sample(X, length(X), replace = T))
  }
  quantile(res, c(1/2 - gamma/2, 1/2 + gamma/2))
}

# we check that it is a valid interval function
superb:::is.errorbar.function("myBootstrapPI.mean")

This is all we need to make the plot which we can compare with the parametric CI

plt1 <- superb(
    score ~ grp, 
    dataFigure1, 
    plotStyle="line",
    statistic = "mean", errorbar = "CI",
    adjustments = list(purpose = "difference")
) + 
xlab("Group") + ylab("Score") + 
labs(title="means and difference-adjusted\n95% confidence intervals") +
coord_cartesian( ylim = c(85,115) ) +
theme_gray(base_size=10) +
scale_x_discrete(labels=c("1" = "Collaborative games", "2" = "Unstructured activity"))

plt2 <- superb(
    score ~ grp, 
    dataFigure1, 
    plotStyle="line",
    statistic = "mean", errorbar = "myBootstrapPI",
    adjustments = list(purpose = "difference")
) + 
xlab("Group") + ylab("Score") + 
labs(title="means and difference-adjusted\n95% bootstrap confidence intervals") +
coord_cartesian( ylim = c(85,115) ) +
theme_gray(base_size=10) +
scale_x_discrete(labels=c("1" = "Collaborative games", "2" = "Unstructured activity"))

library(gridExtra)
plt <- grid.arrange(plt1, plt2, ncol=2)

As seen, there is not much difference between the two. This was expected: when the normality assumption is not invalid, the parametric confidence intervals of the mean (based on this assumption) is identical on average to the bootstrap approach.

In summary

The function superbPlot() is entirely customizable: you can put any descriptive statistic function and any interval function into superbPlot(). In a sense, superbPlot() is simply a proxy that manage the dataset and produces standardized dataframes apt to be transmitted to a ggplot() specification. It is also possible to obtain the summary dataframe by issuing the argument showPlot = FALSE or by using the related function superbData().

The function superbPlot is also customizable with regards to the plot produced. Included in the package are

superbPlot.line : shows the results as points and lines,
superbPlot.point: shows the results as points only,
superbPlot.bar : shows the results using bars,
superbPlot.pointjitter: shows the results with points, and the raw data with jittered points,
superbPlot.pointjitterviolin: also shows violin plot behind the jitter points, and
superbPlot.pointindividualline: show the results with fat points, and individual results with thin lines,
superbPlot.raincloud: show the results along with clouds (violin distributions) and rain drops (jittered raw data),