Prediction fails when both `newdata` and `offset` are provided and model includes `extra_time`.

[sebpardo]

The function predict.sdmTMB() throws an error when both newdata and offset are provided for a model with values for extra_time.

library(sdmTMB)

mesh <- make_mesh(pcod, c("X", "Y"), cutoff = 10)

pcod$os <- rep(log(0.01), nrow(pcod)) # offset

m <- sdmTMB(
  data = pcod,
  formula = density ~ poly(log(depth), 2),
  mesh = mesh,
  offset = pcod$os,
  family = tweedie(link = "log"),
  spatial = "on",
  time = "year",
  extra_time = c(2006, 2008, 2010, 2012, 2014, 2016), 
  spatiotemporal = "AR1"
)

p1 <- predict(m, offset = pcod$os) # This works
p2 <- predict(m, newdata = pcod, offset = pcod$os) # This fails
# Error in `predict()`:
# ! Prediction offset vector does not equal number of rows in prediction dataset.
# Run `rlang::last_trace()` to see where the error occurred.
p3 <- predict(m, newdata = pcod) # This works

Interestingly, you can make it work when "padding" the offset by adding an extra row for each extra_time value:

p4 <- predict(m, newdata = pcod, offset = c(pcod$os, rep(log(0.01), 6))) # This works

Also works when no extra_time is provided:

m2 <- sdmTMB(
  data = pcod,
  formula = density ~ poly(log(depth), 2),
  mesh = mesh,
  offset = pcod$os,
  family = tweedie(link = "log"),
  spatial = "on",
  time = "year",
  spatiotemporal = "AR1"
)

p5 <- predict(m2, newdata = pcod, offset = pcod$os) # This works

I found it informative that the prediction for p1 has six more rows than all other predictions (one for each extra_time), which is especially interesting given that p4 has a longer offset:

dim(p1)
# [1] 2149   19
dim(p3)
# [1] 2143   18
dim(p4)
# [1] 2143   18
dim(p5)
# [1] 2143   18

I think the problematic padding of newdata occurs here and here.

I'm not sure why there is the need to add "fake newdata"(_sdmTMB_fake_nd_ = TRUE) to newdata when there are extra_time values if there are no rows with those timesteps present in newdata? I first thought the easiest solution would be to pad the offset vector internally at the same time as newdata, but then I wondered why you'd want those extra six rows in your prediction output? Not sure what the ideal solution should look like yet.

[seananderson]

Thanks! Fixed. Keeping all time slices in the data at all times makes the bookkeeping simpler within the .cpp.