seananderson
commented
4 years ago I added an example in ?predict.sdmTMB. I think I see how we could automate this type of thing for new data without much trouble, but there remains the problem that you have to be careful about what you actually want to do with the prediction. E.g. if you have + as.factor(year), then what do you want to do in a new year? So forcing the user to do it via the weights argument and think it all through may not be that bad. It also makes sure that everything is set up correctly (e.g. missing years for random walk covariates).
# Forecasting using Eric Ward's hack with the `weights` argument:
# Add on a fake year of data with the year to forecast:
dfake <- rbind(d, d[nrow(d), ])
dfake[nrow(dfake), "year"] <- max(d$year) + 1
tail(dfake) # last row is fake!
weights <- rep(1, nrow(dfake))
weights[length(weights)] <- 0 # set last row weight to 0
dfake$year_factor <- dfake$year
dfake$year_factor[nrow(dfake)] <- max(d$year) # share fixed effect for last 2 years
pcod_spde <- make_spde(dfake$X, dfake$Y, n_knots = 50)
m <- sdmTMB(
data = dfake, formula = density ~ 0 + as.factor(year_factor),
ar1_fields = TRUE, # using an AR1 to have something to forecast with
weights = weights,
include_spatial = TRUE, # could also be `FALSE`
time = "year", spde = pcod_spde, family = tweedie(link = "log"),
silent = FALSE
)
# Add a year to our grid:
qcs_grid$year_factor <- qcs_grid$year
grid2018 <- qcs_grid[qcs_grid$year == 2017L, ]
grid2018$year <- 2018L # `L` because `year` is an integer in the data
qcsgrid_forecast <- rbind(qcs_grid, grid2018)
predictions <- predict(m, newdata = qcsgrid_forecast)
plot_map(predictions, "exp(est)") +
scale_fill_viridis_c(trans = "log10")
plot_map(predictions, "epsilon_st") +
scale_fill_gradient2()
ericward-noaa
Setting weights to zero in the source data for years you want to predict on was a hack suggested by Eric. At the very least, add it in a vignette or example or a note in the help.
Should be doable to let the prediction data frame not match the fitted one, but requires some careful checking of the input data into TMB to make sure it makes sense. Currently it (R) checks to make sure they match on purpose.