epix_slide() converts to a tibble rather than an epi_df - incompatible with arx_forecaster()

[rachlobay]

In the slide function for an epi_archive object, there is a call to convert to a tibble when performing the computation by group:

if (!missing(f)) {
  if (rlang::is_formula(f)) f = rlang::as_function(f)
  
  x = purrr::map_dfr(ref_time_values, function(t) {
    self$as_of(t, min_time_value = t - before_num) %>%
      tibble::as_tibble() %>%  # problem line
      dplyr::group_by(!!!group_by) %>%
      dplyr::group_modify(comp_one_grp,
                          f = f, ..., 
                          time_value = t,
                          key_vars = key_vars,
                          new_col = new_col,
                          .keep = TRUE) %>%
      dplyr::ungroup()
  })
}

So, when a function is used the output from that slide function is a tibble, but the main forecasting function epi_forecaster() from epipredict requires an epi_df to be inputted else it throws an error. Below is an example where we use the latest version of epipredict on GitHub:

library(epipredict)
library(epiprocess)
library(covidcast)
library(data.table)
library(dplyr)
library(tidyr)
library(ggplot2)

y <- covidcast_signals(
  c("doctor-visits", "jhu-csse"),
  c("smoothed_adj_cli", "confirmed_7dav_incidence_prop"),
  start_day = "2020-06-01",
  end_day = "2021-12-01",
  issues = c("2020-06-01", "2021-12-01"),
  geo_type = "state",
  geo_values = c("ca", "fl"))

z <- y[[1]] %>%
  select(geo_value, time_value, version = issue, percent_cli = value) %>%
  as_epi_archive()

z <- epix_merge(
  z, y[[2]] %>%
    select(geo_value, time_value, version = issue, case_rate = value) %>%
    as_epi_archive(), sync = "locf")

fc_time_values <- seq(as.Date("2020-08-01"), as.Date("2021-12-01"),
                      by = "1 month")
ahead = 7

# Old arx_forecaster works fine on the archive
z %>% epix_slide(fc = arx_forecaster(
  percent_cli, case_rate, geo_value, time_value,
  args = arx_args_list(ahead = ahead)),
  n = 120, ref_time_values = fc_time_values)

# New arx does not work on the archive: Error: epi_data must be an `epi_df`.
z %>%
  epix_slide(function(x, ...)
    arx_forecaster(x, outcome = "case_rate",
                       predictors = c("case_rate", "percent_cli"),
                       args = arx_args_list(ahead = ahead))$predictions %>%
      select(-c(geo_value, time_value)),
    n = 120, ref_time_values = fc_time_values, new_col_name = "fc")

So, it looks like we're currently blocked from sliding a forecaster using epi_forecaster() over an epi_archive. Based on a quick inspection, removing the tibble::as_tibble() %>% line seems to do the trick so that epi_forecaster() can work when used as the function in epix_slide(), but the impact of this change should be further investigated before implementing (Some qs about that: Would that change cause any major issues elsewhere? For what purpose was the conversion to tibble done in the first place?). Also, I am not sure that some of the resulting metadata (ex. as_of) makes sense in the outputted epi_df if we do make the above change, but perhaps that could be a separate issue.