|
F-Distribution & Violations of Assumptions
|
# Title: Distribution
# Author: Thomas Verliefde
# Date: 2020/10/19
# Version: 1.1
#
# Log: Fixed issue with unbalanced designs in SampledData (line 393+)
#
# INSTALLS ALL NECESSARY PACKAGES
# list.of.packages = c("shiny","ggplot2","stats","tidyr","dplyr","shinyjs","magrittr",'cowplot','sn');new.packages = list.of.packages[!(list.of.packages %in% installed.packages()[,"Package"])];if(length(new.packages)){install.packages(new.packages,repos="http://cran.us.r-project.org")};lapply(list.of.packages,require,character.only=T);rm(list.of.packages,new.packages)
library(shiny)
library(ggplot2)
library(tidyr)
library(dplyr)
library(shinyjs)
library(magrittr)
library(cowplot)
library(sn)
# Define UI
ui <- fluidPage(
includeScript("../../../Matomo-qhelp.js"),
title = "Distributions",
# Enables the use of shinyjs-package functions (e.g. disabled)
useShinyjs(),
fluidRow(
column(
4,
# Select which distribution to use for the theoretical distributions
selectInput(
'distribution',NULL,
c('Normal Distribution'='Norm','Skew-Normal Distribution'='SN')
),
style="position:relative;
margin-top:10px;
font-size:115%;
font-family:Lucida Console, sans-serif"
),
column(
6,
# The title
HTML('F-Distribution & Violations of Assumptions'),
style="position:relative;
margin-top:8px;
font-family:Lucida Console, sans-serif;
font-weight:500;
font-variant: small-caps;
font-size:200%;"
),
column(
1,
# The sample button.
# Pressing this will resample with the chosen variable values.
actionButton('sample','Sample',icon('refresh'),
style="position:relative;
font-size:115%;
margin-top:10px;
font-family:Lucida Console, sans-serif;
")
)
),
fluidRow(
# The 3 plots displayed
# distplot shows the theoretical plots
# sampleplot shows the sampled data
# fplot shows the calculated F-distribution with theoretical overlay
div(
plotOutput('distplot',width='30%'),
plotOutput('sampleplot',width='30%'),
plotOutput('fplot',width='30%'),
class='shiny-flow-layout',
style='margin-left:15px;
margin-top:10px;'
)
),
fluidRow(
column(
4,
# This changes the appearance of the sliders (.irs elements)
tags$style(type = "text/css", "
.irs {max-height:30px;}
.irs-single {display:none;}
.js-irs-4 .irs-single {display:inline;background:#c8cfa1;color:#000000;}
.js-irs-3 .irs-single {display:inline;background:#c8cfa1;color:#000000;}
.js-irs-2 .irs-single {display:inline;background:#c8cfa1;color:#000000;}
.js-irs-6 .irs-single {display:inline;background:#c8cfa1;color:#000000;}
.irs-grid {display:none !important;}
.irs-bar {background:#c8cfa1;border:1px solid #999fbd;}
.irs-bar-edge {background:#c8cfa1;border-left:1px solid #999fbd;
border-top:1px solid #999fbd;
border-bottom:1px solid #999fbd;}
.irs-min {visibility:hidden !important;
left:-22px;}
.irs-max {visibility:hidden !important;}
.js-irs-4 .irs-min {visibility:visible !important;
left:0px;background:#ffffff;
font-size:80%;}
.js-irs-4 .irs-max {visibility:visible !important;
background:#ffffff;
font-size:80%;}
.js-irs-3 .irs-min {visibility:visible !important;
left:0px;background:#ffffff;
font-size:80%;}
.js-irs-3 .irs-max {visibility:visible !important;
background:#ffffff;
font-size:80%;}
# .js-irs-2 .irs-min {visibility:visible !important;
# left:0px;background:#ffffff;
# font-size:80%;}
# .js-irs-2 .irs-max {visibility:visible !important;
# background:#ffffff;
# font-size:80%;}
.js-irs-5 .irs .irs-max:after {content:'Unbalanced';}
.js-irs-5 .irs .irs-min:after {content:'Balanced';}
.js-irs-0 .irs .irs-max:after {content:'Heterogeneity';}
.js-irs-0 .irs .irs-min:after {content:'Homogeneity';}
.js-irs-1 .irs .irs-max:after {content:'Heteroscedasticity';}
.js-irs-1 .irs .irs-min:after {content:'Homoscedasticity';}
.js-irs-7 .irs .irs-max:after {content:'Dependence';}
.js-irs-7 .irs .irs-min:after {content:'Independence';}
.irs-min:after {visibility: visible !important;}
.irs-max:after {visibility: visible !important;}
.irs-all {top:-10px;}
"),
# Slider indicating group differences on mu/mean/location
sliderInput(
'mu',
HTML("Group Diff: μ"),
min=-1,
max=-.01,
value=-.95,
step=.01
),
# Slider indicating group differences on sigma/variance/scale
sliderInput(
'var',
HTML('Group Diff: σ²'),
min=-1,
max=-.01,
value=-1,
step=.01
),
# Slider indicating the amount of groups
# This slider is not constructed or tested to work with any other value than 4
# As such, this slider is disabeld, and only used as an indicator.
sliderInput(
'groups',
'Groups',
min=2,
max=6,
value=4,
step=1
) %>% disabled # '%>% disabled' makes this slider not clickable
),
column(
4,
# Slider indicating how many times a group of samples is drawn.
# Setting this low will lead to unreliable results.
sliderInput(
'iter',
'Iterations',
min=10,
max=500,
value=250,
step=10
),
# Slider indicating the total sample size of 1 iteration of samples.
sliderInput(
'samplesize',
'Sample Size (n)',
min=16,
max=100,
value=40,
step=4
),
# Slider indicating how balanced the groups are.
# At value = 1 (default), all groups are balanced and have an equal amount of samples.
# At value = 0.01 (minimum), sample sizes differ greatly between groups.
sliderInput(
'balance',
'Group Diff: n',
min=-1,
max=-.01,
value=-1,
step=.01
)
),
column(
4,
# Checkbox to show the non-central F-distribution
# The non-centrality-parameter (lambda) is based on the group difference on mu/location.
# Note that I'm using checkboxGroupInput, instead of checkboxInput.
# The first returns the 'choiceValues', while the latter returns TRUE/FALSE.
# Getting values makes it easier to use the 'switch()' function.
checkboxGroupInput(
'noncentral',NULL,
choiceNames=list(HTML('<b>Show Non-Central F-Dist?</b>')),
choiceValues=list('noncentral')
),
# Slider indicating the significance level
# This slider should technically work with different values.
# But this has not been tested, and was not the aim.
# As such, this slider is disabled, and forms merely an indicator.
sliderInput(
'alpha',
HTML('Significance: α'),
min=0,
max=1,
value=.95,
step=.01
) %>% disabled,
checkboxInput(
'enableDependence',NULL,
label=HTML('<b>Enabling Dependence - Lock Balance</b>'),
value=FALSE
),
# Slider indicating the level of interdependence.
# Not implemented (yet).
sliderInput(
'dependence',
'Dependence',
min=0,
max=1,
value=0,
step=.01
) %>% disabled
)
)
)
# Define server logic
server <- function(input, output, session) {
shinyjs::runjs('toggleCodePosition();')
# Observing the enableDependence checkbox.
# If TRUE (checked), then it resets the balance slider
# If FALSE (unchecked), then it resets the dependence slider
observe(
if(input$enableDependence) {
reset('balance')
} else if(input$enableDependence %>% not) {
reset('dependence')
}
)
# These two lines enable or disable the 'dependence' and 'balance' sliders.
# This is conditional on enableDependence.
# If TRUE (checked), then balance is disabled, and dependence is enabled.
# If FALSE (unchecked), then balance is enabled, and dependence is disabled.
observe(toggleState('dependence', input$enableDependence))
observe(toggleState('balance', input$enableDependence %>% not))
# Reactive switch
# Based on the selected distribution,
# it will select a list of functions to be used.
# Note that currently, this serves little value,
# as for both 'Norm' and 'SN' we are using the same
# functions.
# It could be extended later.
DistFunc=reactive({
switch(input$distribution,
'Norm' = c(dsn,qsn,rsn,df,qf),
'SN' = c(dsn,qsn,rsn,df,qf)
)}
)
# Colour palette that will be used for the different groups.
# One of the reasons the amount of groups only works for 4.
Palette = c('#e66101','#fdb863','#b2abd2','#5e3c99')
# Used in DataDist()
Lims=c(.01,.99)
# sequence of groups, should be c(1,2,3,4), as input$groups is set at 4.
Group=isolate(seq(input$groups))
# MuMin, MuMax are the minimum and maximum values for mu.
# If input$mu (difference between groups on mu) is at abs(0.01),
# 1 group should be at -1, and another group at +1.
MuMin=-1
MuMax=1
# VarMin, VarMax are the minimum and maximum for variance.
# If input$var (difference between groups on var) is at abs(0.01),
# 1 group should be at 1, and another group at 10.
VarMin=1
VarMax=10
# Df1, Df2 determine the degrees of freedom for the F-distributions
Df1=isolate(input$groups-1)
Df2=isolate(input$samplesize-input$groups)
# Non-centrality parameter, changes with group difference on mu (input$mu)
Lambda = reactive(input$samplesize*0.39*(1-abs(input$mu)))
# Skewness factor (called alpha by the sn-distribution functions e.g. sn::dsn)
# Set to 4 if the Skew-normal Distribution is selected, otherwise is at 0.
Skew = reactive({
switch(
input$distribution,
'Norm' = 0,
'SN' = 4
)}
)
# Transformation function for mu, to make the 0.01-1 scale function
MuTrans = function(x,y) {
output = (1-x)*y
return(output)
}
# Transformation function for var, to make the 0.01-1 scale function
VarTrans = function(x,y) {
output = y^(1-x)
return(output)
}
# Transformation function for balance, to make the 0.01-1 scale function
BalTrans = function(x,y,z) {
output = (x / z) %>% {seq((.^y),(.*2 - .^y), length.out = z )} %>% round
return(output)
}
# Create a vector of means, 1 for each group
MuVec = reactive(
sapply(
seq(MuMin,MuMax,along.with=Group),
function(y) MuTrans(abs(input$mu),y)
)
)
# Create a vector of variances, 1 for each group
VarVec = reactive(
sapply(
seq(VarMin,VarMax,along.with=Group),
function(y) VarTrans(abs(input$var),y)
)
)
# Create a vector of sample sizes, 1 for each group
BalVec = reactive(
BalTrans(input$samplesize,abs(input$balance),length(Group))
)
# Find the minimum and maximum values based on the q-plot of the chosen distribution
# This is used as the minimum and maximum values of the x-axis
# for the theoretical plot.
DataDist=reactive({
tibble(
x = sapply(
Group,
function(x) DistFunc()[[2]](Lims,MuVec()[[x]],VarVec()[[x]])
) %>% {c(min(.),max(.))}
)
})
# Create the distplot, showing the theoretical plots.
# There are 4 stat_function plots, 1 for each group.
output$distplot = renderPlot(
ggplot(DataDist(),aes(x=x)) +
sapply(
Group,
function(x) stat_function(
fun=DistFunc()[[1]],n=101,args=list(MuVec()[[x]],VarVec()[[x]],alpha=Skew()),
colour=Palette[x])
) +
labs(x=NULL,y=NULL,title='Theoretical Distributions') +
scale_x_continuous(
breaks=function(x) c(x[1],mean(c(x[1],mean(x))),mean(x),mean(c(x[2],mean(x))),x[2]-.01),
expand=c(0,0)) +
guides(colour=F) +
theme(
panel.grid.major = element_blank(),
panel.grid.minor = element_blank(),
panel.border = element_blank(),
panel.background = element_blank(),
axis.ticks.y = element_blank(),
axis.text.y = element_blank(),
axis.line.y = element_blank(),
axis.ticks.x = element_line(colour='black'),
axis.text.x = element_blank(),
axis.ticks.length = unit(.3,"cm"),
axis.line.x = element_line(colour='black'),
plot.margin = unit(c(0.1,0,.5,0),'cm')
)
)
# This is the system that computes the sampled data.
# It is only reactive on 'input$sample'.
# Normally, something like this should also work with
# eventReactive(), but I didn't get it to work, so I got a workaround.
SampledData =
reactive(
isolate(
replicate(
input$iter,
expr = sapply(
Group,
function(a) replicate(
BalVec()[[a]],
expr = MuVec()[[a]] + DistFunc()[[3]](1,0,VarVec()[[a]])
) %>%
# This part adds a dependence value to all values in a row (over groups).
# Note that if the dependence == 0, this value will always be 0.
# The arbitrary factor 3 is to make sure that the effect is noticable.
as.array %>%
apply(
1,
function(b) {
add(
b,
DistFunc()[[3]](1,0,input$dependence*3)
)
}
)
) %>% t %>% unlist %>% as_tibble %>% unlist
) %>% as_tibble %>% mutate(Grouping = rep(Group,BalVec()) %>% as.factor)
) %T>% {input$sample}
)
# This creates a list of plots showing the sampled data.
# This will be used in the output$sampleplot.
ListPlots = reactive(
lapply(
SampledData() %>% select(1:4),
function(a) {ggplot(SampledData(),aes(x=a,y=Grouping,colour=Grouping)) +
geom_jitter(position = position_jitter(w=0,h=0.1)) +
scale_colour_manual(values=Palette) +
labs(x=NULL,y=NULL)+
guides(colour=F) +
theme(
panel.grid.major = element_blank(),
panel.grid.minor = element_blank(),
panel.border = element_blank(),
panel.background = element_rect(fill='gray98'),
axis.ticks = element_blank(),
axis.text = element_blank(),
axis.line = element_blank()
)}
)
)
# Creates the output object for the sampled plots.
# By utilizing ListPlots, it makes it easier to insert ... after 3.
# This is also the only place (I think) that needs the cowplot package.
output$sampleplot = renderPlot(
plot_grid(
ggdraw() +
draw_label('Sampled Distributions',fontface='bold'),
ListPlots()[[1]],
ListPlots()[[2]],
ListPlots()[[3]],
ggdraw() +
draw_label('. . .',size=30,fontface='bold'),
ListPlots()[[4]],
ncol=1,
align='v',
rel_heights=c(0.2,1,1,1,.3,1)
)
)
# Computing the relevant F-distribution data
FData = reactive(
SampledData() %>%
summarize_at(
vars(-Grouping),
function(x) aov(x ~ Grouping,data=.) %>%
summary %>% unlist(recursive=F) %$% `F value`[[1]]
) %>% gather
)
# FFunc is code for either a central F-distribution, or
# for a noncentral F-distribution, depending on the checkbox earlier.
# The main difference is the addition of ncp=Lambda.
# This would probably also work the other way around, by changing Lambda
# based on the checkbox.
FFunc=reactive({
if(is.null(input$noncentral)) {
list(
"stat_function(
fun=DistFunc()[[4]],n=101,
args=list(df1=Df1,df2=Df2),colour='red1')",
"stat_function(geom='density',
fun=DistFunc()[[4]],n=101,
args=list(df1=Df1,df2=Df2),fill='red1',colour='transparent',
xlim=c(DistFunc()[[5]](input$alpha,Df1,Df2),max(FData()$value)),
alpha=.5)",
"labs(x=NULL,y=NULL,title='F-Distribution')"
)
} else {
list(
"stat_function(
fun=DistFunc()[[4]],n=101,
args=list(df1=Df1,df2=Df2,ncp=Lambda()),colour='red1')",
"stat_function(geom='density',
fun=DistFunc()[[4]],n=101,
args=list(df1=Df1,df2=Df2,ncp=Lambda()),fill='red1',colour='transparent',
xlim=c(DistFunc()[[5]](input$alpha,Df1,Df2,Lambda()),max(FData()$value)),
alpha=.5)",
"labs(x=NULL,y=NULL,title='Non-Central F-Distribution')"
)
}
})
# Creates the third plot, the f-distribution sampled histogram,
# with overlay of the theoretical F-distribution.
# These overlayplots can be found in FFunc()[1] and FFunc()[2].
output$fplot = renderPlot(
ggplot(FData(),aes(x=value)) +
geom_histogram(aes(y=..density..),
bins=50,
fill='transparent',
colour='black') +
eval(parse(text=FFunc()[1])) +
eval(parse(text=FFunc()[2])) +
eval(parse(text=FFunc()[3])) +
scale_x_continuous(
breaks=function(x) c(x[1],mean(c(x[1],mean(x))),mean(x),mean(c(x[2],mean(x))),x[2]-.01),
expand=c(0,0.7)) +
guides(fill=FALSE) +
theme(
panel.grid.major = element_blank(),
panel.grid.minor = element_blank(),
panel.border = element_blank(),
panel.background = element_blank(),
axis.ticks.y = element_blank(),
axis.text.y = element_blank(),
axis.line.y = element_blank(),
axis.ticks.x = element_line(colour='black'),
axis.text.x = element_blank(),
axis.ticks.length = unit(.3,"cm"),
axis.line.x = element_line(colour='black'),
plot.margin = unit(c(0.15,0,.5,0),'cm')
)
)
}
# Run the application
shinyApp(ui = ui, server = server)