Customizing plots

Background

The draw() function combines plots of model terms into a single figure using {patchwork}. Some terms need their own facets or coordinate systems, so separate plots allow {gratia} to display them together.

Use ggplot2 themes to change the appearance of these plots. The & operator applies a theme to every plot in the figure. This article shows how to style {gratia} components and how to build plots from the underlying estimates when you need different layers or mappings from variables to aesthetics.

Adding layers to plots with the `&` operator

We start by simulating some data and fitting a GAM with four smooth functions

library("gratia")
library("mgcv")
#> Loading required package: nlme
#> This is mgcv 1.9-4. For overview type '?mgcv'.
library("ggplot2")
library("dplyr")
#> 
#> Attaching package: 'dplyr'
#> The following object is masked from 'package:nlme':
#> 
#>     collapse
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union
library("patchwork")

# simulate data
n <- 400
eg1 <- data_sim("eg1", n = n, seed = 1)

# fit model
m <- gam(y ~ s(x0) + s(x1) + s(x2) + s(x3),
  data = eg1, method = "REML")

The default plot produced by draw() is

p <- draw(m)
p

If we want to change the theme for the plots, we can’t append a theme() layer to p as this only affects the last plot in the patchwork¹

p + theme_bw()

One way to apply the theme to all plots in the patchwork is to the the & operator.

p & theme_bw()

Styling gratia components

Gratia uses ggplot2 themes for its plot components. Defaults retain the existing appearance. Change an existing figure with & so every panel receives the settings:

p & theme(
  geom.gratia.smooth = element_geom(colour = "navy", linewidth = 0.8),
  geom.gratia.interval = element_geom(fill = "navy"),
  gratia.alpha.interval = 0.15
)

Use geom.gratia.partial.residual for partial residuals, geom.gratia.diagnostic.point for diagnostic points, and geom.gratia.reference for reference lines. Opacity uses numeric settings such as gratia.alpha.interval and gratia.alpha.diagnostic.point. More specific settings, such as geom.gratia.smooth.interval, override the general interval setting. See ?gratia_theme_get for the component list.

Set a package override for subsequent plots without changing other ggplot2 plots. The override is attached when the plot is created:

old_theme <- gratia_theme_set(theme(gratia.alpha.interval = 0.1))
draw(m)

gratia_theme_set(old_theme)

For temporary settings, use withr::with_options() with the option gratia.plot.theme. Settings added to an existing plot follow normal ggplot2 theme merging. Explicit deprecated aesthetic arguments remain fixed overrides, even when they equal the old defaults. For example, replace draw(m, smooth_col = "navy", ci_alpha = 0.15) with the component theme above.

Common replacements for deprecated arguments are:

Previous argument	Theme setting
`smooth_col`	`geom.gratia.smooth = element_geom(colour = ...)`
`ci_col`, `ci_alpha`	`geom.gratia.interval = element_geom(fill = ...)`, `gratia.alpha.interval`
`resid_col`	`geom.gratia.partial.residual = element_geom(colour = ...)`
`contour_col`	`geom.gratia.contour = element_geom(colour = ...)`
`point_col`, `point_alpha` in diagnostics	`geom.gratia.diagnostic.point = element_geom(colour = ...)`, `gratia.alpha.diagnostic.point`
`line_col` in diagnostics	`geom.gratia.diagnostic.reference = element_geom(colour = ...)`
`colour`, `alpha` in smooth samples	`geom.gratia.sample = element_geom(colour = ...)`, `gratia.alpha.sample`
`ci_3d_col`, `ci_3d_alpha`	`geom.gratia.surface.interval = element_geom(fill = ...)`, `gratia.alpha.surface.interval`
`alpha`, `linewidth` in chain diagnostics	`gratia.alpha.chain`, `geom.gratia.chain = element_geom(linewidth = ...)`

For parametric curves use geom.gratia.effect; for derivatives use geom.gratia.derivative. Rootograms have separate rootogram.bar, rootogram.fitted, rootogram.reference and rootogram.warning elements. MRF interval lines use geom.gratia.smooth.interval, independently of the points controlled by geom.gratia.smooth. Diagnostic boxplots use geom.gratia.diagnostic.boxplot.

Optional discrete palettes

Grouped curves and ribbons retain the current Okabe–Ito palette and hue fallback unless another palette is requested. Use the same function for colour and fill to keep groups paired:

tol <- gratia_palette("tol_banded_16")
theme(palette.colour.discrete = tol, palette.fill.discrete = tol)
#> <theme> List of 2
#>  $ palette.colour.discrete:function (n)  
#>  $ palette.fill.discrete  :function (n)  
#>  @ complete: logi FALSE
#>  @ validate: logi TRUE

Paul Tol choices include bright and vibrant (7 colours), muted (9), dark (11), and banded schemes (12, 16 or 20). Dark favours visibility against white; colours within larger banded schemes are less distinct. Finite palettes error when their capacity is exceeded and do not interpolate or recycle colours. See ?gratia_palette for attribution and available names.

Continuous palettes can also be changed through palette.fill.continuous and palette.colour.continuous. Existing scale arguments still accept complete scales when custom transformations, breaks or guides are needed.

Combining individual plots produced by `draw()`

draw() methods like draw.gam() return an object created by patchwork::wrap_plots(), and as a result it isn’t straightforward to combine those objects into a new patchwork

p1 <- draw(m, select = "s(x0)")
p2 <- draw(m, select = "s(x1)")
p3 <- draw(m, select = "s(x2)")
p1 + p2 + p3
#> Error in `wrap_dims()`:
#> ! Need 3 panels, but together `nrow` and `ncol` only provide 1.
#> ℹ Please increase `ncol` and/or `nrow`.

To avoid the error, we need to use patchwork::plot_layout() to set the dimensions we want

p1 + p2 + p3 + plot_layout(ncol = 3)

The above could have been achieved directly via draw()

draw(m, select = c("s(x0)", "s(x1)", "s(x2)"), ncol = 3)

but it is instructive to know how to combine the outputs of draw() should the need arise, such as when you want to create a patchwork of plots from different models.

Drawing surfaces in 3D

Numeric smooth and parametric surfaces can be displayed with surface = "3d". The default remains surface = "raster"; surface = "tile" draws vector rectangles. The 3D display retains the current diverging RdBu palette, fill ranges, theme, and facet styling. Use palette.fill.continuous in theme() to choose another palette.

eg2 <- data_sim("eg2", n = 400, seed = 2)
m_surface <- gam(y ~ te(x, z, k = c(5, 5)) + x:z,
  data = eg2, method = "REML")
draw(m_surface, surface = "3d", n_2d = 20,
  surface_3d = list(yaw = 45, persp = FALSE))

surface_3d is a named list passed to ggcube::coord_3d(). gratia uses only the coordinate system from that function, leaving its additional theme settings out. Options such as yaw, roll, pitch, persp, ratio, and light configure a shared view across all 3D panels. surface_3d = list(dist = 3) controls viewing distance; the top-level dist argument still controls exclusion far from data.

Lighting defaults to ggcube::light(backface_scale = 1), keeping the back faces lit. Supply your own light object to replace this default, for example surface_3d = list(light = ggcube::light(backface_scale = -1)) to restore ggcube’s original back-face shading, or surface_3d = list(light = ggcube::light(method = "none")) to disable lighting. Surface faces have full opacity. Floor observations, boundaries, and contours share the surface’s depth ordering, so hidden parts are occluded by nearer faces.

For a different view for each term, customize the named plots from assemble():

panels <- assemble(m_surface, surface = "3d", n_2d = 20)
# coord_3d() returns a coordinate and a theme; take just the coordinate.
panels[["te(x,z)"]] <- panels[["te(x,z)"]] +
  ggcube::coord_3d(yaw = 45, persp = FALSE,
    light = ggcube::light(backface_scale = 1))[[1]]
#> Coordinate system already present.
#> ℹ Adding new coordinate system, which will replace the existing one.
panels[["x:z"]] <- panels[["x:z"]] +
  ggcube::coord_3d(yaw = -20, persp = FALSE,
    light = ggcube::light(backface_scale = 1))[[1]]
#> Coordinate system already present.
#> ℹ Adding new coordinate system, which will replace the existing one.
wrap_plots(panels)

The two displayed covariates form the horizontal axes, and the partial effect forms the vertical axis. Higher-dimensional smooths and interactions retain conditioning facets; n_2d controls surface resolution and n_3d/n_4d control covariate slices. Contours follow the surface, and available observation locations appear on the floor. No uncertainty surfaces are added. Direct calls to draw(smooth_estimates(model), surface = "3d") and draw(parametric_effects(model), surface = "3d") use the same renderer.

Scattered evaluation data use piecewise-linear Delaunay triangles and require geometry. Support masking and soap-film domain checks also require sf. Faces crossing known excluded regions are discarded conservatively; a coarse mesh may therefore omit additional faces near a boundary. Increase the evaluation density to resolve these areas. Spherical smooths retain their geographic display.

Building your own plot by hand

Themes control the appearance of plots returned by draw(). To choose different layers or map additional variables to aesthetics, build a plot from the estimates returned by {gratia}.

Instead, {gratia} also exports the lower-level functions used draw() so that you can create your own plot using whatever {ggplot2} functions make sense. In the next few code blocks we’ll see how the plot created by draw(m) can be recreated by hand using these lower-level building blocks.

The main thing we need is to evaluate the smooths at values of their covariates. This is done using smooth_estimates(). We also need to add a credible interval to the evaluations, which can be done tidyverse-style via add_confint()

# evaluate the smooths
sm <- smooth_estimates(m) |>
  add_confint()
sm
#> # A tibble: 400 × 11
#>    .smooth .type .by   .estimate   .se     x0    x1    x2    x3 .lower_ci
#>    <chr>   <chr> <chr>     <dbl> <dbl>  <dbl> <dbl> <dbl> <dbl>     <dbl>
#>  1 s(x0)   TPRS  NA       -0.929 0.422 0.0131    NA    NA    NA    -1.76 
#>  2 s(x0)   TPRS  NA       -0.881 0.396 0.0230    NA    NA    NA    -1.66 
#>  3 s(x0)   TPRS  NA       -0.834 0.372 0.0329    NA    NA    NA    -1.56 
#>  4 s(x0)   TPRS  NA       -0.786 0.348 0.0429    NA    NA    NA    -1.47 
#>  5 s(x0)   TPRS  NA       -0.738 0.326 0.0528    NA    NA    NA    -1.38 
#>  6 s(x0)   TPRS  NA       -0.690 0.305 0.0627    NA    NA    NA    -1.29 
#>  7 s(x0)   TPRS  NA       -0.643 0.287 0.0727    NA    NA    NA    -1.20 
#>  8 s(x0)   TPRS  NA       -0.595 0.270 0.0826    NA    NA    NA    -1.12 
#>  9 s(x0)   TPRS  NA       -0.548 0.255 0.0925    NA    NA    NA    -1.05 
#> 10 s(x0)   TPRS  NA       -0.501 0.242 0.102     NA    NA    NA    -0.975
#> # ℹ 390 more rows
#> # ℹ 1 more variable: .upper_ci <dbl>

By default draw.gam() will add partial residuals to the partial effects plots. To achieve the same effect, we need to add the partial residuals to the data used to fit the model. This can be done via add_partial_residuals()

# add partial residuals to data
eg1 <- eg1 |>
  add_partial_residuals(m)

This will² add columns with names "s(x0)". "s(x1)", etc. to the data.

names(eg1)
#>  [1] "y"     "x0"    "x1"    "x2"    "x3"    "f"     "f0"    "f1"    "f2"   
#> [10] "f3"    "s(x0)" "s(x1)" "s(x2)" "s(x3)"

Now we have everything we need to recreate the plots created by draw.gam(). In the code block below we filter sm to focus on a specific smooth, here f(x2) ("s(x2)"), then we add

a rug plot of the observed values of x2,
the credible interval around the estimated smooth,
the partial residuals as a point layer,
the estimated smooth as a line layer,
some annotation

p_sx2 <- sm |>
  filter(.smooth == "s(x2)") |>
  ggplot() +
  geom_rug(aes(x = x2),
    data = eg1,
    sides = "b", length = grid::unit(0.02, "npc")
  ) +
  geom_ribbon(aes(ymin = .lower_ci, ymax = .upper_ci, x = x2),
    alpha = 0.2
  ) +
  geom_point(aes(x = x2, y = `s(x2)`),
    data = eg1, size = 1.5, colour = "steelblue3"
  ) +
  geom_line(aes(x = x2, y = .estimate), linewidth = 1.2) +
  labs(y = "Partial effect", title = "s(x2)")
p_sx2

Assuming we repeat the above steps for the other smooths, creating plot objects p_sx0, p_sx1, p_sx2, and p_sx3 (code not shown), we can complete the plot by creating the patchwork with the desired number of rows and columns

p_sx0 + p_sx1 + p_sx2 + p_sx3 + plot_layout(ncol = 2)

The real benefit of having complete control over how the data are plotted is that you can use the power of {ggplot2} to map additional variables to plot aesthetics. As an example, let’s assume we have a factor variable in the original data and we want to colour the partial residuals according to the levels of this factor. Let’s create that factor

set.seed(12)
eg1 <- eg1 |>
  mutate(fac = sample(letters[1:4], n(), replace = TRUE))

Now we can modify our plotting code to map fac to the colour aesthetic when we plot the partial residuals. To save some typing, we’ll reorder the layers in the plot and add the partial residuals last

plt <- sm |>
  filter(.smooth == "s(x2)") |>
  ggplot() +
  geom_rug(aes(x = x2),
    data = eg1,
    sides = "b", length = grid::unit(0.02, "npc")
  ) +
  geom_ribbon(aes(ymin = .lower_ci, ymax = .upper_ci, x = x2),
    alpha = 0.2
  ) +
  geom_line(aes(x = x2, y = .estimate), linewidth = 1.2) +
  labs(y = "Partial effect", title = "s(x2)")

plt +
  geom_point(
    aes(
      x = x2, y = `s(x2)`,
      colour = fac
    ), # <-- map fac to colour aesthetic
    data = eg1, size = 1.5
  )

We can also do some simple model checking by plotting the smooth with partial residuals coloured according to one of the other covariates (we could also do this by plotting the actual residuals against covariates). In the code chunk below, we map the covariate x1 to both the colour and size aesthetics (note we deleted size = 1.5 to allow the mapping to size)

plt +
  geom_point(
    aes(
      x = x2, y = `s(x2)`,
      colour = x1, size = x1
    ), # <-- map fac to colour aesthetic
    data = eg1, alpha = 0.3
  ) + # <-- deleted cex!!
  scale_colour_viridis_c(option = "plasma")

The resulting plot doesn’t show any particular problems with the model because of the way the data were simulated, but hopefully this illustrates what can be possible once you use the low-level functions provided by {gratia}.