Skip to contents

Correlation heatmaps

Source: vignettes/articles/correlation.Rmd
correlation.Rmd

Correlation heatmaps can be made easily using the ggcorrhm() function. ggcorrhm() is a wrapper for (and shares most of its arguments with) gghm() and calculates the correlation matrix of the input to make a heatmap with gghm(). If one matrix is provided the column-column correlations are plotted. If two matrices are provided, the correlations between the columns in the two matrices are plotted. By using cor_method and cor_use, different method and use arguments can be passed to stats::cor().

ggcorrhm(mtcars)

A heatmap depicting the correlation matrix of the columns in the mtcars dataset. The colour scale goes from -1 to 1 with colours ranging from blue (skyblue2) to white to red (sienna2).

ggcorrhm(iris[1:nrow(mtcars), -5], mtcars)

A heatmap showing the correlation between the columns in the iris data (on the rows) and the columns in mtcars (on the columns).

Changing the colours

ggcorrhm() sets the fill scale to use ggplot2::scale_fill_gradient2() with limits set to [-1, 1]. Unique to ggcorrhm() are the arguments high, mid and low to control the legend colours. Like with gghm(), the bins and limits arguments create a binned scale and change the limits, respectively.

ggcorrhm(mtcars, high = "green", low = "magenta", mid = "yellow", bins = 4L)

A heatmap of the mtcars correlation matrix. The colour scale is changed to consist of 4 bins and the colours go from magenta to green.

If complete control is needed, the col_scale argument can be used to overwrite the default scale. This can be a string specifying a Brewer or Viridis scale, or a ggplot2 scale object.

plt1 <- ggcorrhm(mtcars, col_scale = "RdYlGn")

# Arguments like bins and limits still work with brewer and viridis scales
plt2 <- ggcorrhm(mtcars, col_scale = "PuOr", bins = 5L)

plt1 + plt2

Two correlation heatmaps of mtcars. The first one uses the Brewer RdYlGn palette on a continuous scale, while the second one uses the Brewer PuOr scale with the data put into discrete bins.

Changing the layout

For square heatmaps with identical dimnames (from here on just referred to as square heatmaps), the layout argument can be used to get triangular layouts. The possible layouts are full matrix, top left, top right, bottom left, and bottom right. Note that while the top right and bottom left layouts are just the top right and bottom left triangles of the square matrix, top left and bottom right layouts are actually the bottom left and top right layouts with a flipped y axis to retain the diagonal. This means that some layouts may have the y axis order reversed.

plt1 <- ggcorrhm(mtcars, layout = "topleft")
# The diagonal is displayed by default but can be hidden using the `include_diag` argument
# Layouts can be specified by the first letters e.g. br, tl, f (for full), w (whole, same as full)
plt2 <- ggcorrhm(mtcars, layout = "br", include_diag = FALSE)

plt1 + plt2

Two heatmap of the mtcars correlation matrix. The first plot shows only the top left triangle including the diagonal with names written on it (diagonal running bottom left to top right and row labels reversed compared to the previous heatmaps). The second figure has only the bottom right triangle, but without the diagonal (names are written where the diagonal would be).

Label customisation

For square matrices, the axis names are drawn on the diagonal. The show_names_diag, show_names_rows and show_names_cols arguments control where names should be drawn.

The looks of the names displayed on the diagonal of a square matrix can be adjusted with the names_diag_params argument, which takes a named list of parameters to pass to ggplot2::geom_text(). For names displayed on the x and y axes, use ggplot2::theme() on the generated plot object to change the appearance.

plt1 <- ggcorrhm(mtcars, layout = "bl", legend_order = NA, include_diag = FALSE,
                 names_diag_params = list(
                   angle = 45, vjust = 0.5, hjust = 0.5, colour = "magenta"
                 ))
# Also take vectorised input
set.seed(123)
plt2 <- ggcorrhm(mtcars, layout = "br", include_diag = FALSE,
                 names_diag_params = list(
                   angle = c(0, rep(-45, 9), -90),
                   hjust = 0.7,
                   colour = sample(c("red", "green", "blue"), 11, TRUE)
                 ))

plt1 + plt2

Two mtcars correlation heatmaps with the bottom left and bottom right triangles and no diagonals. In the first plot, the names on the diagonal are at a 45 degree angle and the text is magenta-coloured. The second plot has the names angled 45 degrees in the other direction apart from the names at the top and bottom which are angled at 90 and 0 degrees, respectively. The text colour is picked randomly from red, green, and blue.

Legend position

The legend can be moved around by adding to the theme of the output plot using ggplot2::theme(). Use the legend.position.inside argument to move the legend to the plotting area to make use of the empty space of triangular layouts.

plt1 <- ggcorrhm(mtcars) +
  theme(legend.position = "top",
        legend.title = element_text(vjust = 0.8))
plt2 <- ggcorrhm(mtcars, layout = "br") +
  theme(legend.position = "inside",
        legend.position.inside = c(0.3, 0.75))

plt1 + plt2

Two mtcars correlation heatmaps. In the first one the legend has been moved above the heatmap and is horizontal. In the second the bottom right triangle is drawn and the legend is placed in the empty space where the top left triangle would be drawn.

Clustering and annotation

Just like with gghm(), the heatmap can be clustered and annotated. Triangular layouts limit the positions of annotations and dendrograms to the non-empty sides of the heatmap.

set.seed(123)
# Make a correlation heatmap with a triangular layout, annotations and clustering
row_annot <- data.frame(.names = colnames(mtcars),
                        annot1 = sample(letters[1:3], ncol(mtcars), TRUE),
                        annot2 = rnorm(ncol(mtcars)))
ggcorrhm(mtcars, layout = "tl",
         cluster_rows = TRUE, cluster_cols = TRUE,
         show_dend_rows = FALSE,
         annot_rows_df = row_annot,
         annot_rows_names_side = "top")

A top left triangle correlation heatmap of mtcars. The rows and columns have been clustered and a dendrogram is drawn above the heatmap. On the left side are two thinner columns containing continuous and categorical annotation.

Cell text

The cell_labels argument labels cells with their values or some user-supplied values, explained more in the heatmap article. If cell_labels is a matrix/data frame it should be in the shape of the matrix being plotted (using the same row and column names), which for ggcorrhm is the correlation matrix. To keep in mind for triangular layouts is that the top left or bottom right triangles will display the bottom left or top right triangles with the y axis order flipped, and this is reflected in the cell labels too.

lab_mat <- matrix(1:121, nrow = 11, ncol = 11, byrow = TRUE,
                  dimnames = list(colnames(mtcars), colnames(mtcars)))

plt_list <- lapply(c("tl", "tr", "bl", "br"), function(lt) {
  # Play a bit with the legend for the last plot
  legend_order <- if (lt == "br") 1 else NA
  ggcorrhm(mtcars, layout = lt, legend_order = legend_order,
           cell_labels = lab_mat, show_names_diag = FALSE) +
    theme(legend.position = "inside",
          legend.position.inside = c(0.025, 1.075))
})

wrap_plots(plt_list)

Four triangulra mtcars correlation heatmaps arranged in a 2x2 grid (top left triangle in the top left, top right in the top right etc) so that a diamond-shaped gap is formed in the middle. The cells are labelled from 1 to 121, filling the matrix rowwise. The top left and bottom left heatmaps are the same but with the y-axes flipped. The same relationship is true for the top right and bottom right heatmaps. In the middle (in the gap between the four plots) is one legend.

Cell shape

Correlation heatmaps are sometimes plotted with circles. Passing a value from 1 to 25 to the mode argument allows for different cell shapes. Shapes 21-25 support filling and use fill scales while 1-20 are not filled and use colour scales (relevant if a scale object is passed). The shape size is set to vary between the values in the size_range argument (not found in gghm()), scaling with the absolute value of the correlation. size_scale can be provided to overwrite this behaviour. The size legend is hidden by default when using ggcorrhm() as it would only show the absolute values of the correlations (but the legend can be shown with the legend_order argument if necessary).

Setting mode to ‘text’ writes the values in empty cells, with the text colour scaling with the correlation values.

# If mode is a number, the R pch symbols are used, meaning 21 produces filled circles.
plt1 <- ggcorrhm(mtcars, mode = 21) + labs(title = "mode = 21")
# It is also possible to use shapes other than circles. Only 21-25 support filling.
# Change the size range using the size_range argument
plt2 <- ggcorrhm(mtcars, mode = 23, size_range = c(2, 7)) +
  # Can add a grid in the background with ggplot2::theme()
  theme(panel.grid.major = element_line(colour = "grey90")) +
  labs(title = "mode = 23")

plt1 + plt2

Two mtcars correlation heatmaps. The first one has circles instead of cells. The colour scales with the correlation and the size with the absolute value of the correlation. The second plot has squares (diamonds) and a major grid is added in the background.

It’s also possible to draw mixed layouts by providing two values to layout and mode. See the mixed layouts article for more details.

# Default modes for mixed layouts is 'heatmap' and 'text'
ggcorrhm(mtcars, layout = c("tr", "bl"))

mtcars heatmap where the top right triangle (and the diagonal) has filled cells and the bottom left triangle has cells containing text instead. The text shows the correlation values and the colour scales with the correlation like the cells in the other triangle.

P-values

ggcorrhm() adds support for correlation p-values which can be calculated by setting the p_values argument to TRUE. The values can be adjusted for multiple testing by passing a string specifying the adjustment method to the p_adjust argument. If the correlation matrix is square, the diagonal is excluded and only half of the off-diagonal values are used for the adjustment. The calculated p-values will be included in the data if return_data is TRUE. P-values and adjusted p-values for the correlations of the diagonal in a square matrix are set to 0 as the correlations are always 1.

By default, cells are marked with asterisks if p_values is TRUE. The p_thresholds argument controls the p-value thresholds. p_thresholds must be NULL (for no thresholds or symbols) or a named numeric vector where the values specify the thresholds (in ascending order) and the names are the symbols to use when an adjusted p-value is below each threshold. stats::symnum() is used to convert p-values and the last value of p_thresholds must be 1 (or any higher number) to set the upper bound. By default, p_thresholds is set to be c("***" = 0.001, "**" = 0.01, "*" = 0.05", 1). As can be seen, 1 is left unnamed so that any p-values between 0.05 and 1 are unmarked.

# P-value symbols added to the plot
plt1 <- ggcorrhm(mtcars, p_values = TRUE, p_adjust = "bonferroni")
# Changing the symbols
plt2 <- ggcorrhm(mtcars, p_values = TRUE, p_adjust = "bonferroni",
                 p_thresholds = c("+" = 0.01, "*" = 0.05, "ns" = 1))

plt1 + plt2

Two mtcars correlation heatmaps. The cells are coloured by correlation and contain text that depends on the p-values of the correlation coefficients. In the first one cells are marked with up to three asterisks. The second plot instead has a mix of cells with a plus symbol, an asterisk, and 'ns' written in them.

In a mixed layout it is possible to display p-values in only half of the plot by passing a vector or list of length two to the p_values argument. If cell_labels is TRUE (so that correlation values are written in the cells), the symbols from p_thresholds are added at the end of the correlation values.

# P-values only in the bottom right, mode 'none' for labels with empty background
ggcorrhm(mtcars, layout = c("tl", "br"), mode = c("hm", "none"), p_adjust = "bonferroni",
         p_values = c(FALSE, TRUE), cell_labels = c(FALSE, TRUE))

A mtcars correlation heatmaps with filled cells in the top left and labelled white cells in the bottom right. There are correlation values written in the bottom right, with asterisks signifying the p-value.

It is also possible to write out the p-values instead of correlation values by setting both cell_labels and cell_label_p to TRUE.

# P-values and labels only in one half, correlation labels in the other
ggcorrhm(mtcars, layout = c("tl", "br"), mode = c("hm", "none"), p_adjust = "bonferroni",
         p_values = c(FALSE, TRUE), cell_labels = TRUE,
         cell_label_p = c(FALSE, TRUE))

A mtcars correlation heatmaps with filled cells in the top left and labelled white cells in the bottom right. There are labels in both triangles; correlation values in the top left and p-values with asterisks in the bottom right.

This is also carried over to ‘text’ mode.

# Same as previous but text mode
ggcorrhm(mtcars, layout = c("tl", "br"), mode = c("hm", "text"), p_adjust = "bonferroni",
         p_values = c(FALSE, TRUE),
         # In text mode the labels are displayed regardless so cell_labels can be FALSE
         cell_labels = c(TRUE, FALSE),
         cell_label_p = c(FALSE, TRUE))

A mtcars correlation heatmaps with filled cells in the top left and labelled white cells in the bottom right. There are labels in both triangles; correlation values in the top left and p-values with asterisks in the bottom right. The p-value labels are coloured with the same scale as the filled cells in the top left triangle (scaling with the correlation values).

By returning the data and using ggplot2 functions it is possible to be more creative with the p-value marking than ggcorrhm() allows out of the box.

p <- ggcorrhm(mtcars, p_values = TRUE, p_thresholds = NULL,
              p_adjust = "bonferroni", return_data = TRUE,
              layout = "br", include_diag = FALSE)
p$plot_data <- mutate(p$plot_data,
                      p_lty = case_when(p_adj < 0.05 ~ 1, TRUE ~ 3),
                      p_lwd = case_when(p_adj < 0.05 ~ 1, TRUE ~ 0.5),
                      p_size = -log10(p_adj))
ggcorrhm(mtcars, border_lty = p$plot_data$p_lty, border_lwd = p$plot_data$p_lwd,
         layout = "br", include_diag = FALSE) +
  geom_point(aes(x = col, y = row, size = p_size), data = p$plot_data %>%
               filter(p_adj < 0.05, as.character(row) != as.character(col)),
             shape = "*") +
  scale_size(range = c(4, 8)) +
  labs(size = "-log10(p_adj)")

mtcars correlation heatmap showing the bottom right triangle. Some cells are marked with asterisks whose sizes scale with -log10 of the adjusted p-values. Those cells also have thick grey border. Other cells have thin, dotted borders.

p <- ggcorrhm(mtcars, p_values = TRUE, p_thresholds = NULL,
              p_adjust = "bonferroni", return_data = TRUE,
              layout = "tl", include_diag = FALSE)
ggcorrhm(mtcars, layout = "tl", include_diag = FALSE) +
  geom_point(data = p$plot_data %>%
               # Mark cells above a threshold
               filter(p_adj >= 0.05, as.character(row) != as.character(col)),
             mapping = aes(x = col, y = row),
             shape = 4, size = 8)

A top left correlation heatmap of mtcars. Some cells are marked with a large x (indicating that the adjusted p-value is above 0.05).

Some extra plots

Triangular layouts require square matrices, but not symmetric ones (even if correlation heatmaps typically are symmetric). This means that the different triangles could show different information, such as the results of different correlation methods.

# Use different correlation methods
pear_cor <- cor(mtcars, method = "pearson")
kend_cor <- cor(mtcars, method = "kendall")
# Combine
cor_df <- pear_cor
cor_df[lower.tri(cor_df)] <- kend_cor[lower.tri(kend_cor)]

plt1 <- ggcorrhm(cor_df, cor_in = TRUE, cell_labels = TRUE,
                 col_name = "Correlation") +
  # Add axis titles
  labs(x = "Kendall tau", y = "Pearson r") +
  theme(axis.title = element_text())

# With different colour scales, although that makes it hard to compare
plt2 <- gghm(cor_df, layout = c("bl", "tr"), mode = c("hm", "hm"),
             col_scale = c("RdBu_rev", "PRGn_rev"),
             col_name = c("Pearson r", "Kendall tau"),
             split_diag = TRUE, limits = c(-1, 1))

plt1 + plt2

Two correlation heatmaps of the mtcars data, with the bottom left triangle showing Pearson correlation and the top right triangle showing Kendall correlation. The first heatmap uses a single colour scale (blue to white to red) for the whole heatmap and has cells labelled with correlation values. The second heatmap uses two different colour scales for the different triangles (blue-white-red for bottom left, green-white-purple for top right). The cells in the diagonal are drawn as triangles, splitting the diagonal into two in the middle.

Other packages

A benefit of making the plot with ggplot2 is the possibility of modifying and arranging the plot using ggplot2 and its many extensions. ggcorrheatmap is convenient for making simple correlation heatmaps, with options for extensive customisation of the appearance and layout, and built-in support for clustering (with dendrograms) and annotation. Another package for making correlation heatmaps with ggplot2 is ggcorrplot which is faster and has more ways of visualising p-values. If having the plot made with ggplot2 is not necessary, the corrplot package can make flexible correlation plots with many alternative visualisations.