## tidyverse components
library(dplyr)    ## data manipulation
library(tidyr)    ## ditto (pivot_longer)
library(readr)    ## read CSV files
library(ggplot2); theme_set(theme_bw())
## other
library(pgmm)     ## for olives data
library(agridat)  ## agricultural data sets
library(skimr)    ## for text-based summaries
library(GGally)   ## pairs plots etc.
library(corrplot) ## correlation plots

• method = "loess" (see ?loess, Wikipedia)
  • local polynomial regression (quadratic by default)
  • used by default for N<1000
  • from ?geom_smooth: “Somewhat anecdotally, ‘loess’ gives a better appearance, but is O(N^2) in memory, so does not work for larger datasets”
  • increase span (in method.args) to get smoother result
  • could use method.args = list(degree=1) to get locally linear fit …

• method="gam"
• generalized additive model
• penalized regression (decides on smoothness automatically)
• typically results in smoother curve than loess
• may oversmooth (try adjusting max knots, e.g. formula = y ~ s(x,k=10,bs="cs") for a cubic spline with 10 knots)
  • formulas within geom_smooth expressed in terms of x and y (regardless of original var names)
• can use e.g. method.args = list(family=poisson) to account for data ranges/variance

• generalized linear models: method="lm" or method="glm"
• simplest, smoothest
• also most constraining/parametric
• can fit polynomials, e.g. formula= y ~ poly(x,2)
• use method.arg=list(family=...) as appropriate

• “melt” data to long format (pivot_longer())
• minor abuse of faceting

cars_long <- pivot_longer(mtcars,
                          -mpg,  ## DON'T melt mpg
                          names_to="xvar", values_to="xval")
g1 <- (ggplot(cars_long,aes(xval,mpg)) +
       geom_point() +
       facet_wrap(~xvar, scale="free_x") +
       ## include defaults to suppress message
       geom_smooth(method = "loess", formula = y ~ x)
)

print(g1)

data(olive, package="pgmm")
data(fisher.barley, package="agridat")

library("sos")
findFn("barley")      ## full-text search of CRAN
?RSiteSearch          ## similar
help.search("barley",agrep=FALSE) ## search in *installed* package

ff <- (fisher.barley
  %>% mutate(gen = reorder(gen,yield),
             env = reorder(env,yield),
             year = factor(year))
  %>% arrange(gen, env) ## order of rows (shouldn't affect computations)
)

gg0 <- (ggplot(ff,
               aes(x=env, y=yield, colour=year, group=year))
  ## geom_boxplot(aes(x=gen,y=yield,fill=factor(year)))
  + geom_point()+ geom_line()
  + facet_wrap(~gen,nrow=1)  ## 1-row layout
)

print(gg0)

• Hard to see x-axis labels: try rotating? coord_flip() doesn’t interact well with faceting (see also ggstance package)
• colours not optimal (almost always use colours from colorspace or scale_colour_brewer() or …) but probably fine
• optimal aspect ratio?
• squash spaces between panels (Tufte advice)?

gg0_h <- (ggplot(ff,
               aes(y=env, x=yield, colour=year, group=year))
  + geom_point()
  + geom_path() ## geom_line() doesn't work when we flip x/y
  + facet_wrap(~gen, nrow=1)
  + theme(panel.spacing = grid::unit(0,"lines")) ## squash spaces
  + coord_fixed(ratio=20) ## aspect ratio
)

print(gg0_h)

maybe use ncol=1, coord_fixed(ratio=1/20)? Then yields would be aligned along the same axis (but a tall skinny plot might not be convenient)

Could calculate differences …

## restore names to data set 
olive_regions <-
    read_csv("../data/olive_regions.csv")
olive2 <- 
  (full_join(olive_regions, olive,  ## region/area names
             by = c("region_num" = "Region",
           "area_num" = "Area"))
    %>% select(-c(region_num, area_num))
    %>% mutate(across(c("region", "area"), factor))
  )
colvec <- c("red", "blue", "goldenrod")

Setup: define a function to draw density plots by group

panel.density <- function(x, ...) {
  usr <- par("usr"); on.exit(par(usr))  ## restore user coordinates
  par(usr = c(usr[1:2], 0, 1.5) )       ## modify y limits
  h <- lapply(split(x, olive2$region), density)
  for (i in 1:length(h)) {
    lines(h[[i]]$x, h[[i]]$y / max(h[[i]]$y), col = colvec[i])
  }
}
num_cols <- 3:ncol(olive2)

pairs(olive2[ , num_cols ],
      gap = 0,     ## no spaces
      cex = 0.5,   ## smaller points
      col = colvec[olive2$region],
      diag.panel = panel.density)

print(ggp1)

• To make everything in a graph smaller/larger, either
  • set the base_size= argument (in pointsize) in theme_bw() (or whatever)
  • make the output size larger (in ggsave or fig.width/fig.height args to an RMarkdown chunk)

car::scatterplotMatrix(olive2[,num_cols],
                       groups=olive2$region,
                       gap=0)

What about just showing correlations? (Probably not “exploratory”.)

library(corrplot)
corrplot(cor(olive2[,num_cols]),method="ellipse")

corrplot.mixed(cor(olive2[,num_cols]),lower="number",upper="ellipse")