For a paper I’ve been writing I was running linear mixed effects models with categorical fixed effects to compare the amount of pine weevil damage present in different forest sites. To evaluate my models, I wanted to run pairwise comparisons between sites in the model to see which sites were significantly different. I decided to run a Tukey’s HSD (Honestly Significant Difference) test on Estimated Marginal Means of sites in the model. I’m doing my analysis in R, so all the code is R code.
I used emmeans::emmeans()
to calculate the EMMs, which returns a special object of class emmGrid
. emmeans::pwpp()
is a handy function to create and plot the results from a Tukey’s HSD (or other test method) pairwise comparison of the EMMs, with P-value along the x axis and categories along the y axis. Lines connect pairwise categories and are placed along the x axis to denote the significance of their difference.
The basic plot is fine, and is certainly useful for interpretation, but I want to include the plots in my paper and they look a bit too much like the ggplot()
default. The plotting method also makes it difficult to customise the plot. The basic code to produce the plot is below:
# Packages
library(glmmTMB)
library(emmeans)
library(ggplot2)
# Import data
df <- read.csv("~/Desktop/pwpp_data.csv")
# Run mixed effects model
mod <- glmmTMB(mm2_damage ~ site_code + (1|family),
data = df)
# Estimate Marginal Means
tukey <- emmeans(mod, "site_code")
# Run Pairwise compari
pwpp_results <- pwpp(tukey, values = TRUE, sort = FALSE)
# Look at basic plot
pwpp_results
The plot looks like this:
As the object pwpp_results
is a glorified ggplot()
object, it’s possible to extract the data used to make the plot and store it in tidy dataframes:
# Extract data from plot object
marg_vals <- data.frame(
y = pwpp_results$layers[[3]]$data$site_code,
label = pwpp_results$layers[[3]]$data$fmtval)
p_vals <- data.frame(
x = pwpp_results$data$p.value,
plus = pwpp_results$data$plus,
minus = pwpp_results$data$minus,
midpt = pwpp_results$data$midpt)
marg_vals
holds data on the marginal values for each category, which appear as labels down the left hand side of the plot. p_vals
contains data for drawing the lines.
I also wanted to create a colour palette for my customised plot, which I made here, at Colorgorical :
# Create colour palette
site_pal <- c("#270fe2", "#75ae0a", "#b427b7", "#14e54b",
"#8e4380", "#0b5313", "#d992e2", "#7ba979",
"#fc2c44", "#1ce0b2", "#900e08", "#37bad7")
I can then create the plot:
pwpp_ggplot <- ggplot() +
geom_segment(data = p_vals,
aes(x = x, xend = x, y = plus, yend = midpt, colour = minus)) +
geom_point(data = p_vals,
aes(x = x, y = plus, colour = minus),
size = 3) +
geom_label(data = marg_vals,
aes(x = 0.01, y = y, label = label),
label.padding = unit(0.15, "lines"), hjust = "right") +
geom_vline(aes(xintercept = 0.05),
linetype = 2) +
ylab("Site") +
xlab("Tukey-adjusted P value") +
theme_classic() +
theme(panel.grid.major.y = element_line(colour="#E0E0E0"),
axis.title = element_text(size = 14),
axis.text = element_text(size = 12),
axis.text.y = element_text(colour = site_pal),
legend.position = "none") +
scale_x_continuous(breaks = c(0, 0.05, 0.1, 0.5, 1)) +
scale_colour_manual(values = site_pal) +
coord_trans(x = "log10", clip = "off")
geom_segment
draws the lines. Each line is actually composed of two line segments which meet at p_vals$midpt
and are coloured according to the opposite point.
geom_label
plots the marginal values, which are placed at x = 0.01
and are right aligned so they are unlikely to overlap a comparison line.
geom_vline
denotes the p = 0.05 significance line.
panel.grid.major.y
draws lines for each category, which helps when reading the graph and matching line ends with categories.
scale_x_continuous
marks breaks for a number of common significance thresholds.
coord_trans
log transforms the x axis so that lines near the low end of the P value scale are more spaced out, as they are more important for interpretation, clip = "off"
ensures that the full range of P values is shown, so plots are interpretable across models with different categories.
The final plot looks like this: