Does stan_gamm4 work with a basis other than the default?

[jgellar]

Summary:

Weird-looking results when using stan_gamm4 with a non-default basis.

Description:

I have recently discovered the stan_gamm4 function to allow for GAMs. I'd like to fit some models with different bases, however I've tried some that give weird results. See below.

Reproducible Steps:

library(dplyr)
library(tidyr)
library(ggplot2)
library(rstanarm)
options(mc.cores = parallel::detectCores())

# Simulate some data
N <- 1000
dat <- data.frame(x = rnorm(N)) %>%
  mutate(y = sin(pi*x) + rnorm(N, sd = .5))
x0 <- seq(min(dat$x), max(dat$x), by = .1)

# Fit a model with default basis
fit1 <- stan_gamm4(y ~ s(x), data = dat)
plot1 <- plot_nonlinear(fit1)
plot1 +
  geom_point(aes(x,y), data = dat, inherit.aes = FALSE, alpha = .25, size = .25) +
  geom_line(aes(x0, y0), data = data.frame(x0 = x0, y0 = sin(pi*x0)),
            inherit.aes = FALSE, colour = "red")

This plot looks great. Note that the red line is the "truth", and the black dots are the observed points:

However, when we change the basis...

# Fit a model with a p-spline basis of order 8
fit2 <- stan_gamm4(y ~ s(x, bs = "ps", k = 8), data = dat)
plot2 <- plot_nonlinear(fit2)
plot2

Now this doesn't look right. In particular, what's with the range of x values (from about -0.2 to 0.4), which is far tighter than the range of simulated x values (closer to -2 to 2)? I don't think this is just a problem with plot_nonlinear(), because the following also gives a weird result:

yhat <- posterior_predict(fit_rsag4_2, newdata = data.frame(x = x0)) %>% colMeans()
plot(x0, yhat, type = "l")

Here, the estimate isn't even smooth... it's really bumpy. What's going on?

RStanARM Version:

2.13.1

R Version:

MRO 3.3.1

Operating System:

Windows 10

[bgoodri]

We don't change the basis (at least not on purpose) but there is a bug in the prediction code with newdata following stan_gamm4: #132 . Basically, we are not accounting for the centering and scaling done by the mgcv package..

[bgoodri]

The underlying bug was fixed, so this should be fine now. Please install the master branch from GitHub.