dropping rows with NA's in relevant columns only

[dsweber2]

The reason our forecasters were underperforming was that when merging the chng and hhs dataset, we were removing any rows with NA's. This fixes things to only remove NA's inside a forecaster, and only for columns actually used in forecasting. For an example of the problematic data, compare

hhs_archive_alone <- tar_read("hhs_archive_data_2022") %>%
          select(geo_value, time_value, value, issue) %>%
          rename("hhs" := value) %>%
          rename(version = issue) %>%
          as_epi_archive(
            geo_type = "state",
            time_type = "day",
            compactify = TRUE
          )
max(hhs_archive_alone$as_of(as.Date("2021-11-29"))$time_value)

which is the last hhs data, while

chng_archive_alone <- tar_read("chng_archive_data_2022") %>%
          select(geo_value, time_value, value, issue) %>%
          rename("chng" := value) %>%
          rename(version = issue) %>%
          as_epi_archive(
            geo_type = "state",
            time_type = "day",
            compactify = TRUE
          )
max(chng_archive_alone$as_of(as.Date("2021-11-29"))$time_value)

which is the last chng data

chng_archive_alone$as_of(as.Date("2021-11-29")) %>% arrange(time_value) %>% tail
both <- tar_read(joined_archive_data_2022)
max(both$as_of(as.Date("2021-11-29"))$time_value)

joined matches chng:

both <- tar_read(joined_archive_data_2022)
max(both$as_of(as.Date("2021-11-29"))$time_value)

closes #93

[dsweber2]

I'm going to run just the data targets in production to make sure this does the right thing too.

[dsweber2]

so after renv::install("targets@1.3.2", "tarchetypes@0.7.9"); renv::snapshot(), this does seem to be working.

I also added the necessary changes to get DEBUG_MODE=true to actually get browser() working. Doesn't seem to be a way around having to run tar_make(name, callr_function=NULL).

[dsweber2]

We definitely need to figure out a better way to deal with dates for slide_forecaster, though I'm not sure what exactly.