Search code examples
rggplot2pcaggplotlyggbiplot

ggplot2: Adjusting label positions for the PCA loadings in a PCA biplot in R

Issue:

I have produced a PCA biplot using the packages ggbiplot/ggplot2. I have lengthened the loadings (arrows) using the function geom_segment() and I would like to delete the original loadings (short arrows), keep the longer loadings (new arrows), keep the labels with a grey background, but reposition them so they are non-overlapping and sit nicely at the end of the longer loading arrowheads. On the PCA biplot below, there are ten loadings but there are only eight parameters. Something is not right.

I have researched extensively through StackOverflow, on the web, and I've asked the R Studio Community to resolve my issue, although, the only information that I can find is either through different biplot functions or a reference to other entirely different packages for PCA (MASS, factoextra, FactoMineR, PCAtools, and many others), none of which address the question that I would like to answer.

I would really like to continue using ggbiplot/ggplot2 to get a better understanding of both packages and I prefer the visual representation of the PCA biplot (see below) I've created. I would ideally like the biplot to look like the desired output (see below). Please see the data attached below.

If anyone can help, I would be deeply appreciative.

Many thanks in advance

R-code:

install.packages("remotes")
remotes::install_github("vqv/ggbiplot")
install_github("vqv/ggbiplot")

#install.packages("devtools")
library(devtools)
library(ggbiplot)
library(ggplot2)
library(remotes)

#You can do a PCA to visualize the difference between the groups using the standardised box cox data
PCA=prcomp(Whistle_Parameters[2:18], center = TRUE, scale=TRUE, retx = T)

#PCA loadings
PCAloadings <- data.frame(Variables = rownames(PCA$rotation), PCA$rotation)

#Place label into an object
Country<-data$Country

#PCA biplots
PCA_plot1<-ggbiplot::ggbiplot(PCA, ellipse=TRUE, circle=TRUE, varname.adjust = 2.5, groups=Country, var.scale = 1) +
                     ggtitle("PCA of Acoustic Parameters") +
                     theme(plot.title = element_text(hjust = 0.5)) +
                     theme_minimal() +
                     theme(panel.background = element_blank(), 
                     panel.grid.major = element_blank(), 
                     panel.grid.minor = element_blank(),
                     panel.border = element_blank()) + 
                     geom_vline(xintercept = 0, lty = 2) +
                     geom_hline(yintercept = 0, lty = 2) +
                     theme(axis.line.x = element_line(color="black", size = 0.8),
                     axis.line.y = element_line(color="black", size = 0.8)) +
                     geom_segment(data = PCAloadings, aes(x = 0, y = 0, xend = (PC1*4.6),
                                yend = (PC2*4.6)), arrow = arrow(length = unit(1/2, "picas")),
                                color = "black", alpha=0.75) +
                     scale_color_manual(values=c('#E69F00', '#56B4E9')) 

#The options for styling the plot within the function itself are somewhat limited, but since it produces a 
#ggplot object, we can re-specify the necessary layers. The following code should work on any object 
#output from ggbiplot. First we find the geom segment and geom text layers:
seg <- which(sapply(PCA_plot1$layers, function(x) class(x$geom)[1] == 'GeomSegment'))
txt <- which(sapply(PCA_plot1$layers, function(x) class(x$geom)[1] == 'GeomText'))

#We can change the colour and width of the segments by doing
PCA_plot1$layers[[seg[1]]]$aes_params$colour <- 'black' 
PCA_plot1$layers[[seg[2]]]$aes_params$colour <- 'black'

#To change the labels to have a gray background, we need to overwrite the geom_text layer with a geom_label layer:
PCA_plot1$layers[[txt]] <- geom_label(aes(x = xvar, y = yvar, label = PCAloadings$Variables,
                                         angle = 0.45, hjust = 0.5, fontface = "bold"), 
                                         label.size = NA,
                                         data = PCA_plot1$layers[[txt]]$data, 
                                         fill = '#dddddd80')

PCA_plot1

PCA Biplot

enter image description here

Desired output

enter image description here

Data

structure(list(Country = structure(c(1L, 1L, 1L, 1L, 1L, 1L, 
1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L, 
1L, 1L, 1L, 1L, 1L, 1L, 1L, 1L), .Label = c("France", "Germany"
), class = "factor"), Low.Freq = c(1229.098358827, 759.408910773596, 
627.156561121131, 857.914227798394, 816.020512657709, 726.252107003186, 
603.388640229802, 1120.43591201848, 521.916239230762, 66.4277073927702, 
842.827028732445, 1548.00685289626, 743.586603639966, 982.298813187027, 
938.898554484786, 908.067281510105, 1650.51840217509, 435.837013213028, 
905.64518713548, 621.693057238002, 408.874626715846, 763.284854056395, 
1163.36397892984, 1267.90558781017, 1561.05494859439, 269.110242829792, 
1628.20258277437, 1381.52801863709, 1259.64885050619, 582.429604337893
), High.Freq = c(270.037998321385, 673.100410969792, 1354.51641087434, 
582.818682820139, 1949.42791374982, 533.072062804075, 1462.73353623344, 
1475.85981044777, 1672.72713391206, 1360.85064740235, 1027.62671423916, 
1637.72929840934, 555.708652550379, 683.537132648398, 1714.01010661954, 
267.117743854174, 738.883902818488, 842.919932827166, 124.511854388999, 
1940.70836004547, 991.37814311059, 1959.73951887933, 435.882938574683, 
223.944759894009, 827.050231552967, 1929.9835959516, 731.983627515309, 
934.515637669084, 1381.80407878684, 1735.12129509753), Start.Freq = c(209.223178720873, 
1243.93824398519, 714.942866646311, 1230.88587487336, 1133.38920481274, 
655.640254812419, 176.783487591076, 566.793710992312, 1259.4101411541, 
135.19626803044, 1188.65745695622, 1055.06564740433, 145.269654935287, 
994.102001940972, 611.97046714505, 1239.25416627405, 273.254811174704, 
1187.0983873612, 532.369927415851, 730.144132713145, 306.959091815357, 
761.432150933258, 833.35964575595, 633.492439842001, 1011.91529244509, 
1002.6837164403, 874.223664731894, 1039.77329580107, 571.716041690428, 
358.734914494325), End.Freq = c(3100.50977989246, 2865.99128764993, 
3749.07057886566, 1662.75251781181, 3469.5453928947, 1223.29004528624, 
1571.69393815622, 3877.90064918956, 3467.9046701139, 2812.46639335828, 
3344.48917919081, 823.479192696172, 3587.72640978872, 1943.42597579601, 
3726.46938122543, 904.270327650973, 2199.36865174236, 1608.75061469279, 
1233.86988042306, 3573.54022639883, 3840.13648049746, 2696.92512488242, 
2152.58952962537, 2225.74334558365, 2836.8576276391, 3909.86535579565, 
2642.3514330105, 3830.39875611625, 2532.59589574087, 2047.68204963624
), Peak.Freq = c(615.103200058515, 781.386010343022, 1254.22250479111, 
1042.32081012698, 1785.4136370848, 609.196990586287, 936.529532621528, 
628.617072934145, 1138.00887772997, 965.752651960148, 331.381776986669, 
831.243197072226, 1396.09323352817, 796.57855885715, 1434.02692184993, 
581.91826512844, 1482.84787412806, 712.229765737013, 711.849861782499, 
379.432018940052, 1495.87429192735, 1375.38825516007, 1568.51147252198, 
39.6849748542959, 254.973241980045, 526.048000326837, 1599.41223732841, 
1723.40465012645, 461.157566614546, 847.403323972557), Center.Freq = c(-0.00396318509300687, 
0.377462792184857, 1.66350671788962, -0.152573382048654, 0.438259482923988, 
1.62149800844459, 0.501892326424285, -0.166579179714419, 1.06081611813746, 
1.39199162769052, 1.52164843383928, -0.389958351497529, 0.00261034688899059, 
0.0726410215179534, 1.00473421813784, 1.27072495569536, 1.41569796343226, 
0.737375815997266, 0.412628778604207, 0.51099123600198, 1.65512836540775, 
-1.12408230668747, 0.438260531725931, -1.11347230908714, 1.09021071848368, 
1.26465014876586, -0.663254496003035, 0.64384027394782, 1.29816899903361, 
0.0302328674903059), Delta.Freq = c(2374.48934930825, 2535.28648042237, 
930.363518659463, 2372.94461226817, 2578.50041236941, 1652.93682378145, 
2412.64071270543, 1643.35808756239, 1597.6988634255, 2347.87731769764, 
1545.35983248752, 417.894712991398, 676.404759114593, 2717.74464723351, 
2750.52013318133, 1387.50061490775, 1088.18301844773, 208.885548316239, 
982.856603814324, 1304.55461743298, 2064.83914948351, 1454.17493801179, 
1975.72909682146, 1340.40119652782, 1358.81720189322, 398.974468430338, 
1807.83210129773, 197.995771350184, 1458.91300578134, 2459.54002342707
), Delta.Time = c(1.52332103330495, -0.729369599299347, 0.5446606158259, 
-0.0806278952890181, -1.03355982391612, 0.381391555011319, -0.710006011318096, 
0.184876103317229, -0.0939796220798944, 0.878826387745255, 0.889598364118577, 
0.929698941247702, 0.734996499853458, -0.43364546563554, -0.176575903721404, 
0.556057576098353, -0.31543237357059, 1.31950129257089, 1.08676447814548, 
-1.08756351145615, -0.163851619861579, -0.945982375537661, 0.473134073749239, 
-0.231569591521918, -0.565159893817776, 1.14721196081124, -1.14555651287826, 
1.60486934195338, -1.00704726744845, 1.14020903183312)), row.names = c(NA, 
30L), class = "data.frame")
Solution

  • How about just saving the loadings in a separate dataframe and then plotting with ggplot::geom_segment? Like this:

    #You can do a PCA to visualize the difference between the groups using the standardised box cox data
PCA = prcomp(df[2:9], center = TRUE, scale=TRUE, retx = T)

#PCA loadings
PCAloadings <- data.frame(Variables = rownames(PCA$rotation), PCA$rotation)

# add PCA scores to the dataset
df[, c('PC1', 'PC2')] = PCA$x[, 1:2]

# save variable loadings in a separate dataset
rot = as.data.frame(PCA$rotation[, 1:2])
rot$var = rownames(PCA$rotation)

# rescale the loadings to fit nicely within the scatterplot of our data
mult = max(abs(df[, c('PC1', 'PC2')])) / max(abs(rot[, 1:2])) / 2
rot[, 1:2] = rot[, 1:2] * mult

# ggplot the scatterplot and rotation taken from separate data.frames
ggplot(data = rot, aes(x = 0, y = 0, xend = PC1, yend = PC2, label = var)) +
  geom_point(data = df, aes(PC1, PC2), inherit.aes = FALSE) +
  geom_segment(color = 'red', arrow = arrow(length = unit(0.03, "npc"))) +
  geom_label(aes(PC1 * 1.2, PC2 * 1.2)) +
  theme_bw() +
  theme(panel.grid = element_blank())

# if there are many variables to plot, you can play with ggrepel 
library(ggrepel)  
ggplot(data = rot, aes(x = 0, y = 0, xend = PC1, yend = PC2, label = var)) +
  geom_point(data = df, aes(PC1, PC2), inherit.aes = FALSE) +
  geom_segment(color = 'red', arrow = arrow(length = unit(0.03, "npc"))) +
  geom_label_repel(aes(PC1 * 1.2, PC2 * 1.2)) +
  theme_bw() +
  theme(panel.grid = element_blank())