Plotting using gratia::draw

[rvaradhan]

I fit a GAM model: gam(y ~ trt, s(time, by=trt, method="REML"), where the response y is measured at times time. The variable trt is a factor with 4 levels. Using "gratia" I would like to plot the fitted GAM model along with the original data so that I can visually inspect the adequacy of the fit. There are 4 panels in my plot corresponding to each level of "trt". How can I add the observed data points to each panel, in addition to the fitted smoother?

[rvaradhan]

I would like to plot on the original scale of the response rather than on the partial effect scale.

[gavinsimpson]

What you're looking for is to predict from the model. This is usually done via predict(), but gratia::fitted_values() can will facilitate this:

library("mgcv")
library("gratia")
library("ggplot2")

eg_data <- data_sim("eg4", n = 400, dist = "normal", scale = 2, seed = 1)
m <- gam(y ~ fac + s(x2, by = fac), data = eg_data, method="REML", family = gaussian())

ds <- data_slice(m, x2 = evenly(x2), fac = evenly(fac))
fv <- fitted_values(m)

fv |>
  ggplot(aes(x = x2, y = .fitted, group = fac)) +
  geom_ribbon(aes(ymin = .lower_ci, ymax = .upper_ci, fill = fac), alpha = 0.2) +
  geom_line(aes(colour = fac)) +
  geom_point(data = eg_data, aes(colour = fac, y = y)) +
  facet_wrap(~ fac)

[rvaradhan]

Dear Gavin - this is terrific. I was able to get beautiful plots! Thank you very much.

[rvaradhan]

Gavin - Is it possible to extend your solution to GAMM object? I tried this but it doesn't work. Thanks for any suggestion.

[rvaradhan]

Sorry - it works fine for the GAMM object as well.

[kobby184]

I am plotting a multinomial GAM , one of my smooth terms is s(longitude, latitude). values seem to be on top of each other. Is there a way around it. Thank you

[kobby184]

I use draw(model, rug=FALSE)

[gavinsimpson]

It's probably just overplotting individual data points that are observed at the same location. I could reduce these to distinct values of the covariate when plotting. But usually, people either want this on or off.

[gavinsimpson]

In the development version, draw(model) will now only show the unique data values by default. See #327