A function to describe revision behavior for an archive.

revision_summary removes all missing values (if requested), and then computes some basic statistics about the revision behavior of an archive, returning a tibble summarizing the revisions per time_value+epi_key features. If print_inform is true, it prints a concise summary. The columns returned are:

n_revisions: the total number of revisions for that entry
min_lag: the minimum time to any value (if drop_nas=FALSE, this includes NA's)
max_lag: the amount of time until the final (new) version (same caveat for drop_nas=FALSE, though it is far less likely to matter)
min_value: the minimum value across revisions
max_value: the maximum value across revisions
median_value: the median value across revisions
spread: the difference between the smallest and largest values (this always excludes NA values)
rel_spread: spread divided by the largest value (so it will always be less than 1). Note that this need not be the final value. It will be NA whenever spread is 0.
time_near_latest: the time taken for the revisions to settle to within within_latest (default 20%) of the final value and stay there. For example, consider the series (0, 20, 99, 150, 102, 100); then time_near_latest is 5, since even though 99 is within 20%, it is outside the window afterwards at 150.

Usage

revision_summary(
  epi_arch,
  ...,
  drop_nas = TRUE,
  print_inform = TRUE,
  min_waiting_period = as.difftime(60, units = "days"),
  within_latest = 0.2,
  quick_revision = as.difftime(3, units = "days"),
  few_revisions = 3,
  abs_spread_threshold = NULL,
  rel_spread_threshold = 0.1,
  compactify_tol = .Machine$double.eps^0.5,
  should_compactify = TRUE
)

Arguments

epi_arch: an epi_archive to be analyzed
...: <tidyselect>, used to choose the column to summarize. If empty, it chooses the first. Currently only implemented for one column at a time.
drop_nas: bool, drop any NA values from the archive? After dropping NA's compactify is run again to make sure there are no duplicate values from occasions when the signal is revised to NA, and then back to its immediately-preceding value.
print_inform: bool, determines whether to print summary information, or only return the full summary tibble
min_waiting_period: difftime, integer or NULL. Sets a cutoff: any time_values not earlier than min_waiting_period before versions_end are removed. min_waiting_period should characterize the typical time during which revisions occur. The default of 60 days corresponds to a typical final value for case counts as reported in the context of insurance. To avoid this filtering, either set to NULL or 0.
within_latest: double between 0 and 1. Determines the threshold used for the time_to
quick_revision: difftime or integer (integer is treated as days), for the printed summary, the amount of time between the final revision and the actual time_value to consider the revision quickly resolved. Default of 3 days
few_revisions: integer, for the printed summary, the upper bound on the number of revisions to consider "few". Default is 3.
abs_spread_threshold: numeric, for the printed summary, the maximum spread used to characterize revisions which don't actually change very much. Default is 5% of the maximum value in the dataset, but this is the most unit dependent of values, and likely needs to be chosen appropriate for the scale of the dataset.
rel_spread_threshold: float between 0 and 1, for the printed summary, the relative spread fraction used to characterize revisions which don't actually change very much. Default is .1, or 10% of the final value
compactify_tol: float, used if drop_nas=TRUE, it determines the threshold for when two floats are considered identical.
should_compactify: bool. Compactify if TRUE.

Examples

revision_example <- revision_summary(archive_cases_dv_subset, percent_cli)
#> Min lag (time to first version):
#>      min median     mean     max
#>   3 days 3 days 3.5 days 12 days
#> Fraction of epi_key+time_values with
#> No revisions:
#> • 0 out of 1,956 (0%)
#> Quick revisions (last revision within 3 days of the `time_value`):
#> • 0 out of 1,956 (0%)
#> Few revisions (At most 3 revisions for that `time_value`):
#> • 0 out of 1,956 (0%)
#> 
#> Fraction of revised epi_key+time_values which have:
#> Less than 0.1 spread in relative value:
#> • 91 out of 1,956 (4.65%)
#> Spread of more than 2.22056495 in actual value (when revised):
#> • 671 out of 1,956 (34.3%)
#> days until within 20% of the latest value:
#>      min median     mean     max
#>   3 days 5 days 9.1 days 67 days
revision_example %>% arrange(desc(spread))
#> # A tibble: 1,956 × 11
#>    time_value geo_value n_revisions min_lag max_lag time_near_latest spread
#>    <date>     <chr>           <dbl> <drtn>  <drtn>  <drtn>            <dbl>
#>  1 2020-12-26 ca                 62 3 days  73 days  6 days           14.1 
#>  2 2020-12-25 ca                 62 3 days  73 days  7 days           13.2 
#>  3 2020-11-27 fl                 66 3 days  73 days  4 days           12.0 
#>  4 2021-09-27 fl                 43 3 days  63 days 59 days            9.79
#>  5 2020-12-25 fl                 62 3 days  73 days  4 days            9.75
#>  6 2021-09-26 fl                 43 4 days  64 days 60 days            9.48
#>  7 2021-09-27 ca                 43 3 days  63 days  8 days            9.31
#>  8 2021-09-25 fl                 43 5 days  65 days 61 days            8.83
#>  9 2020-11-05 ny                 66 3 days  73 days 11 days            8.64
#> 10 2020-11-27 tx                 66 3 days  73 days 10 days            8.56
#> # ℹ 1,946 more rows
#> # ℹ 4 more variables: rel_spread <dbl>, min_value <dbl>, max_value <dbl>,
#> #   median_value <dbl>