posterior_linpred() does not use newdata for stann_gamm4 models

[noamross]

Summary:

posterior_linpred() does not seem to use newdata for stann_gamm4 models

Description:

When I pass a new data frame to posterior_linpred() along with a model fit with stan_gamm4, the result is identical to the case where I don't pass any new data at all.

Reproducible Steps:

library(rstanarm)
library(tidyverse)
gtemp <- read.delim("http://www.metoffice.gov.uk/hadobs/hadcrut4/data/current/time_series/HadCRUT.4.5.0.0.annual_ns_avg.txt",
                    sep="", header = FALSE) %>%
  select(1,2) %>%
  set_names(c("Year", "Temperature")) %>%
  mutate(Year = as.double(Year))

st_mod2 <- stan_gamm4(Temperature ~ s(Year), data = gtemp)
preds2 <- posterior_linpred(st_mod2)
preds2_d <- posterior_linpred(st_mod2, newdata=data.frame(Year=(1850:2016) + 0.1))
identical(preds2, preds2_d)

This results in TRUE

RStanARM Version:

2.13.1 (installed from github)

R Version:

3.3.1

Operating System:

OS X 10.12.1

[bgoodri]

I believe the issue here has to do with the fact that you are adding a constant (0.1). If you add runif(nrow(gtemp)), then the linear predictors are not identical. I'm not sure that we have the right behavior here, but when mgcv constructs s(Year) it is centered.

[bgoodri]

So, I guess we should tack on the original mean(s) multiplied by the coefficients.

[noamross]

Yeah, I think you need to correct for mgcv's centering if you're reconstructing the gam object when doing this. For my particular application, I need to add a constant in order to estimate the derivatives of the smooths, but in any case we'll want to be able to get predictions on an absolute scale.

[bgoodri]

This should be fixed on GitHub now. It goes back to the mgcv-style parameterization, so now Stan is calling predict.gam with type = "lpmatrix".