`detect_outlr_stl()` requires gap filling while `detect_outlr_rm()` does not? Document this.

[rachlobay]

It seems like detect_outlr_stl() requires gap filling. For example, if we manually remove two rows from the below epi_df & pop that into detect_outlr_stl(), then we inevitably get an error about the data containing implicit gaps in time. In contrast, if we use the same data in detect_outlr_rm(), there are no complaints from that function (likely due to epi_slide). So, there should be strong documentation that explains how missing rows are handled by each function. As well, we should probably update the epi_slide vignette to explain how it handles missing rows of data (which is not likely to be uncommon for users of this package).

Ex. showing error in detect_outlr_stl() and no problem in detect_outlr_rm()

library(epidatr)
library(epiprocess)
library(dplyr)
library(tidyr)

# Load #s of new confirmed COVID-19 cases, daily, for FL
# over a fairly large time window
x <- covidcast(
  data_source = "jhu-csse",
  signals = "confirmed_incidence_num",
  time_type = "day",
  geo_type = "state",
  time_values = epirange(20200601, 20220601),
  geo_values = "fl",
  as_of = 20220606
) %>%
  fetch_tbl() %>%
  select(geo_value, time_value, cases = value) %>%
  as_epi_df()

x<- x[-c(2,10),] # Remove some rows from x

y = x$cases
x = x$time_value
# The below should all be the default values from detect_outlr_stl()
n_trend = 21
n_seasonal = 21
n_threshold = 21
seasonal_period = NULL
log_transform = FALSE
detect_negatives = FALSE
detection_multiplier = 2.5
min_radius = 0
replacement_multiplier = 0

# Below is the first part of the detect_outlr_stl() function 
  # Transform if requested
  if (log_transform) {
    # Replace all negative values with 0
    y = pmax(0, y)
    offset = as.integer(any(y == 0))
    y = log(y + offset)
  }
  
  # Make a tsibble for fabletools, setup and run STL
  z_tsibble = tsibble::tsibble(x = x, y = y, index = x)
  
  stl_formula = y ~ trend(window = n_trend) +
    season(period = seasonal_period, window = n_seasonal)
  
  stl_components = z_tsibble %>%
    fabletools::model(feasts::STL(stl_formula, robust = TRUE)) %>%
    generics::components() %>%
    tibble::as_tibble() %>%
    dplyr::select(trend:remainder) %>%
    dplyr::rename_with(~ "seasonal", tidyselect::starts_with("season")) %>% 
    dplyr::rename(resid = remainder)


# Now, the same data when inputted into detect_outlr_rm() has no apparent problem

  x <- covidcast(
    data_source = "jhu-csse",
    signals = "confirmed_incidence_num",
    time_type = "day",
    geo_type = "state",
    time_values = epirange(20200601, 20220601),
    geo_values = "fl",
    as_of = 20220606
  ) %>%
    fetch_tbl() %>%
    select(geo_value, time_value, cases = value) %>%
    as_epi_df()
  
  x<- x[-c(2,10),] # Remove some rows from x
  
  x <- x %>%
    group_by(geo_value) %>%
    mutate(outlier_info  = detect_outlr_rm(
      x = time_value, y = cases),
      detection_multiplier = 2.5) %>% #%% change this to something larger potentially or nah?
    unnest(outlier_info)
  
  x

[brookslogan]

I'm marking this as P3 because it might be sort of working as intended/stated as is, and because we probably want to prioritize #256, which would change whether/how we approach this one. It'd be nice if all detect_outlr* and epi_cor methods could have the same approach here, so we don't need to describe these details for every method.