superb

Source: R/superb.R

The function superb() plots standard error or confidence interval for various descriptive statistics under various designs, sampling schemes, population size and purposes, according to the superb framework. See (Cousineau et al. 2021) for more. The functions superb() is now the entry point to realize summary plots. Compared to the previously documented superbPlot(), superb() is based on formula and accept long and wide format.

superb(
  formula,
  data,
  WSFactors = NULL,
  WSDesign = NULL,
  factorOrder = NULL,
  statistic = "mean",
  errorbar = "CI",
  gamma = 0.95,
  adjustments = list(purpose = "single", popSize = Inf, decorrelation = "none",
    samplingDesign = "SRS"),
  showPlot = TRUE,
  plotStyle = "bar",
  preprocessfct = NULL,
  postprocessfct = NULL,
  clusterColumn = NULL,
  ...
)

Arguments

formula

a formula describing the design of the data frame

data

Dataframe in wide or long format

WSFactors

The name of the within-subject factor(s)

WSDesign

the within-subject design if not a full factorial design (default "fullfactorial")

factorOrder

Order of factors as shown in the graph (in that order: x axis, groups, horizontal panels, vertical panels)

statistic

The summary statistic function to use as a string

errorbar

The function that computes the error bar. Should be "CI" or "SE" or any function name if you defined a custom function. Default to "CI"

gamma

The coverage factor; necessary when errorbar == "CI". Default is 0.95.

adjustments

List of adjustments as described below. Default is adjustments = list(purpose = "single", popSize = Inf, decorrelation = "none", samplingDesign = "SRS")

showPlot

Defaults to TRUE. Set to FALSE if you want the output to be the summary statistics and intervals.

plotStyle

The type of object to plot on the graph. See full list below. Defaults to "bar".

preprocessfct

is a transform (or vector of) to be performed first on data matrix of each group

postprocessfct

is a transform (or vector of)

clusterColumn

used in conjunction with samplingDesign = "CRS", indicates which column contains the cluster membership

...

In addition to the parameters above, superbPlot also accept a number of optional arguments that will be transmitted to the plotting function, such as pointParams (a list of ggplot2 parameters to input inside geoms; see ?geom_bar2) and errorbarParams (a list of ggplot2 parameters for geom_errorbar; see ?geom_errorbar)

Value

a plot with the correct error bars or a table of those summary statistics. The plot is a ggplot2 object with can be modified with additional declarations.

Details

The possible adjustements are the following

  • popsize: Size of the population under study. Defaults to Inf

  • purpose: The purpose of the comparisons. Defaults to "single". Can be "single", "difference", or "tryon".

  • decorrelation: Decorrelation method for repeated measure designs. Chooses among the methods "CM", "LM", "CA", "UA", "LDr" (with r an integer) or "none". Defaults to "none". "CA" is correlation-adjusted (Cousineau 2019) ; "CM" is Cousineau-Morey (Baguley 2012) ; "LM" is Loftus and Masson (Loftus and Masson 1994) ; "UA" is based on the unitary Alpha method (derived from the Cronbach alpha; see (Laurencelle and Cousineau 2023) ). "LDr" is local decorrelation (useful for long time series with autoregressive correlation structures; see (Cousineau et al. 2024) ).

  • samplingDesign: Sampling method to obtain the sample. implemented sampling is "SRS" (Simple Randomize Sampling) and "CRS" (Cluster-Randomized Sampling).

The formulas can be for long format data using | notation, e.g.,

  • superb( extra ~ group | ID, sleep )

or for wide format, using cbind() or crange() notation, e.g.,

  • superb( cbind(DV.1.1, DV.2.1,DV.1.2, DV.2.2,DV.1.3, DV.2.3) ~ . , dta, WSFactors = c("a(2)","b(3)"))

  • superb( crange(DV.1.1, DV.2.3) ~ . , dta, WSFactors = c("a(2)","b(3)"))

The layouts for plots are the following:

  • These are basic plots:

    • "bar" Shows the summary statistics with bars and error bars;

    • "line" Shows the summary statistics with lines connecting the conditions over the first factor;

    • "point" Shows the summary statistics with isolated points

    • "lineband" illustrates the confidence intervals as a band;

  • These plots add distributional information in addition

    • "pointjitter" Shows the summary statistics along with jittered points depicting the raw data;

    • "pointjitterviolin" Also adds violin plots to the previous layout

    • "pointindividualline" Connects the raw data with line along the first factor (which should be a repeated-measure factor)

    • "raincloud" Illustrates the distribution with a cloud (half_violin_plot) and jittered dots next to it. Looks better when coordinates are flipped +coord_flip()

    • "corset" illustrates within-subject designs with individual lines and clouds.

  • Circular plots (aka radar plots) results from the following layouts:

    • "circularpoint" Shows the summary statistics with isolated points

    • "circularline" Shows the summary statistics with lines;

    • "circularlineband" Also adds error bands instead of error bars;

    • "circularpointjitter" Shows summary statistics and error bars but also jittered dots;

    • "circularpointlinejitter" Same as previous layout, but connect the points with lines. New layouts are added from times to time. Personalized layouts can also be created (see Vignette5).

References

Baguley T (2012). “Calculating and graphing within-subject confidence intervals for ANOVA.” Behavior Research Methods, 44, 158 -- 175. doi:10.3758/s13428-011-0123-7 .

Cousineau D (2019). “Correlation-adjusted standard errors and confidence intervals for within-subject designs: A simple multiplicative approach.” The Quantitative Methods for Psychology, 15, 226 -- 241. doi:10.20982/tqmp.15.3.p226 .

Cousineau D, Goulet M, Harding B (2021). “Summary plots with adjusted error bars: The superb framework with an implementation in R.” Advances in Methods and Practices in Psychological Science, 4, 1--18. doi:10.1177/25152459211035109 .

Cousineau D, Proulx A, Potvin-Pilon A, Fiset D (2024). “Local decorrelation for error bars in time series.” The Quantitative Methods for Psychology, 20(2), 173-185. doi:10.20982/tqmp.20.2.p173 .

Laurencelle L, Cousineau D (2023). “Analysis of proportions using arcsine transform with any experimental design.” Frontiers in Psychology, 13, 1045436. doi:10.3389/fpsyg.2022.1045436 .

Loftus GR, Masson MEJ (1994). “Using confidence intervals in within-subject designs.” Psychonomic Bulletin & Review, 1, 476 -- 490. doi:10.3758/BF03210951 .

Examples

######################################################################

# Basic example using a built-in dataframe as data. 
# By default, the mean is computed and the error bar are 95% confidence intervals
superb(len ~ dose + supp, ToothGrowth) 


# Example changing the summary statistics to the median and
# the error bar to 80% confidence intervals
superb(len ~ dose + supp, ToothGrowth,
       statistic = "median", errorbar = "CI", gamma = .80) 


# Example introducing adjustments for pairwise comparisons 
# and assuming that the whole population is limited to 200 persons
superb(len ~ dose + supp, ToothGrowth,  
  adjustments = list( purpose = "difference", popSize = 200) )


# This example adds ggplot directives to the plot produced
library(ggplot2)
superb(len ~ dose + supp, ToothGrowth) + 
xlab("Dose") + ylab("Tooth Growth") +
theme_bw()


######################################################################

# The following examples are based on repeated measures
library(gridExtra)
options(superb.feedback = 'none') # shut down 'warnings' and 'design' interpretation messages

# A simple example: The sleep data
# The sleep data are paired data showing the additional time of sleep with 
# the soporific drug #1 (("group = 1") and with the soporific drug #2 ("group = 2"). 
# There is 10 participants with two measurements.
# sleep is available in long format 

# Makes the plots first without decorrelation:
superb( extra ~ group | ID, sleep )

# As seen the error bar are very long. Lets take into consideration correlation...
# ...  with decorrelation (technique Correlation-adjusted CA):
superb(extra ~ group | ID, sleep, 
  # only difference:
  adjustments = list(purpose = "difference", decorrelation = "CA")
)

# The error bars shortened as the correlation is substantial (r = .795).


######################################################################

# Another example: The Orange data
# This example contains 5 trees whose diameter (in mm) has been measured at various age (in days):
data(Orange)

# Makes the plots first without decorrelation:
p1 <- superb( circumference ~ age | Tree, Orange,
  adjustments = list(purpose = "difference", decorrelation = "none")
) + 
  xlab("Age level") + ylab("Trunk diameter (mm)") +
  coord_cartesian( ylim = c(0,250) ) + labs(title="''Standalone'' confidence intervals")
# ... and then with decorrelation (technique Correlation-adjusted CA):
p2 <- superb( circumference ~ age | Tree, Orange,
  adjustments = list(purpose = "difference", decorrelation = "CA")
) + 
  xlab("Age level") + ylab("Trunk diameter (mm)") +
  coord_cartesian( ylim = c(0,250) ) + labs(title="Decorrelated confidence intervals")

# You can present both plots side-by-side
grid.arrange(p1, p2, ncol=2)


######################################################################