Using extra_time to predict/extrapolate over future years

[seananderson]

Question by email:

I will want to use the output of my model to interpolate over future years. If my understanding is right, this is what the argument extra_time allows one to do. However, I do not see any example with this argument. Do you have any examples or cues about what I have to specify for this argument in the function (TRUE/FALSE, on?, specify years for which I want interpolation,….)? In that case, 'year’ has also have to be specify for the "time" argument (eg time='year')?

[...] I would also like to take account for the fact there they may have some autocorrelation between successive dates during which surveys were conducted. So I want I want to include date as a random effect in the model. I think it is possible to do it using the time_varying or spatiotemporal argument. Am I right? If yes, is there any conflict if I’m using a different variable for time and time_varying?

[seananderson]

Sorry, I didn’t realize there weren’t examples in there right now with extra_time. There’s a set of slides here: https://pbs-assess.github.io/sdmTMB-teaching/noaa-psaw-2022/ (see forecasting) That's from this repository.

And a vignette Julia is developing based on that that I haven’t integrated yet:
https://github.com/pbs-assess/sdmTMB/blob/main/scratch/julia-vignettes/forecasting.Rmd

extra_time should be a vector of numeric values representing the missing or extra time slices. E.g. extra_time = c(2023, 2024)

[...] I would also like to take account for the fact there they may have some autocorrelation between successive dates during which surveys were conducted. So I want I want to include date as a random effect in the model. I think it is possible to do it using the time_varying or spatiotemporal argument. Am I right? If yes, is there any conflict if I’m using a different variable for time and time_varying?

There are many ways, and I think we cover that in the slides above.

Assuming you mean ‘year’ for an annual survey, year could be a random walk:

time = “year”,
time_varying = ~ 1

The random walk is by whatever the time argument is defined as, so time and time_varying are linked.

or an IID random intercept:

dat$fyear <- as.factor(dat$year)
formula = … + (1 | fyear),
time = “year”,

or a smoother:

formula = s(year),
time = “year”,

or absorbed fully into the random walk or AR1 random field:

time = “year”,
spatiotemporal = “rw”

or

spatiotemporal = “ar1”

If you mean day-of-year or something, that’s probably best entered as a smoother.