baselinenowcast 0.2.2

Package

• Add a deprecation tag to the package data germany_covid19_hosp and make new package data named covid19_data which extends through August 2022.
• Fixed argument passing in baselinenowcast.reporting_triangle() and baselinenowcast.data.frame to estimate_uncertainty() and sample_nowcasts().
• Added a more informative error message when the user specifies a number of reference dates in estimate_delay() that is larger than the number of rows in the reporting triangle.
• Reduced test run time for CRAN by using 10 draws instead of the default 1000 in the slow ref_time_aggregator tests, skipping those tests on CRAN with skip_on_cran(), and setting a seed in a stochastic test that could fail intermittently (#488).

Documentation

• Moved the evaluation example vignette to a website-only pkgdown article in vignettes/articles/ so that it is no longer built on CRAN. The article is rendered when the pkgdown site is built in CI (#488).

baselinenowcast 0.2.1 (GitHub-only release)

Package

• Add as_forecast_sample() and as_forecast_point() S3 methods for baselinenowcast_df objects, enabling direct conversion of probabilistic or point nowcasts to scoringutils forecast objects for scoring against the latest observed counts (#418). baselinenowcast_df objects now carry a logical nowcast column flagging the right-truncated reference dates that were actually nowcast (as opposed to fully observed), and the converters use it to score only those dates.
• Place every internal check, assert, and validate function next to its consumer (class invariants in R/reporting_triangle-class.R and R/baselinenowcast_df-class.R; per-caller helpers in the file that calls them) and remove R/validate.R. Generalise assert_rep_tri_class() with an arg_name argument and use it to replace duplicated inline class checks in truncate_to_quantile(), truncate_to_delay(), and the reporting triangle getters and methods (#300).
• Add converters between reporting_triangle and reviser vintages formats via as_reviser_vintages() and as_reporting_triangle.tbl_pubdate(), enabling use of reviser’s vintage analysis and state space nowcasting methods alongside baselinenowcast’s nowcasting functionality (#429).
• Add truncate_to_date(), a date-based wrapper around truncate_to_row() that drops rows with reference dates after a given cutoff (#447).

Documentation

• Modified NSSP nowcast vignette to do vectorised string matching instead of row-wise and to simplify the pre-processing to avoid a pivot to a wide form dataframe to speed-up performance.
• Added a new vignette which provides an example of how to evaluate visually and quantitatively a single nowcast model specification, as well as how to use quantitative evaluation metrics to choose a model specification.

baselinenowcast 0.2.0

This is the first stable release providing core functionality for producing probabilistic or point nowcasts from a dataframe with incident case counts indexed by reference dates and report dates. It supports doing so across multiple strata, either through creating reporting_triangle objects for each strata and nowcasting independently, or nowcasting directly from a dataframe containing multiple strata. The reporting_triangle object allows users to specify the unit of the delay, and provides helpful print and summary S3 methods, as well as functionality to truncate to a specific delay or a percentile of the observed delays. A paper validating and evaluating the performance of the methods is available at https://wellcomeopenresearch.org/articles/10-614/v1.

Package

• fill_triangle() has been removed. Use estimate_and_apply_delay() instead, which now supports optional delay estimation via the delay_pmf parameter (#334).
• fill_triangles() has been removed. Use estimate_and_apply_delays() instead (#334).
• Rename truncate_triangle() to truncate_to_row() and truncate_triangles() to truncate_to_rows() to clarify that these functions truncate by row count and distinguish them from other truncation utilities such as truncate_to_quantile() and truncate_to_delay() (#336).
• Use the as_reporting_triangle and baselinenowcast workflow in the vignette which walks through a nowcasting example applied to syndromic surveillance data in the U.S.
• Add preprocess parameter to estimate_delay() to control handling of negative values in reporting triangles. Set to preprocess_negative_values by default to redistribute negative values to earlier delays, or set to NULL to preserve negative PMF entries. This allows the method to work with reporting corrections that result in net downward adjustments at specific delays (#278).
• Export preprocess_negative_values() function to allow users to manually handle negative values in reporting triangles by redistributing them to earlier delays (#278).
• Add a baselinenowcast.data.frame() method which ingests a data.frame with one or more strata to nowcast and returns a baselinenowcast_df object.
• Add converters between reporting_triangle and ChainLadder triangle formats via as_ChainLadder_triangle() and as_reporting_triangle.triangle(), enabling use of ChainLadder’s mature claims reserving methods alongside baselinenowcast’s nowcasting functionality.
• Add a baselinenowcast.reporting_triangle() method which ingests a reporting_triangle object and produces a baselinenowcast_df object which is a data.frame with nowcasts and associated metadata.
• Add a as_reporting_triangle() S3 method which ingests a data.frame or matrix and returns a reporting_triangle object, which will be used as an input to the eventual baselinenowcast() function.
• Add a vignette which demonstrates how to pre-process for nowcasting syndromic surveillance system data using the U.S. National Syndromic Surveillance Program (NSSP) data as a case study, and then use baselinenowcast to nowcast cases of a specific syndromic surveillance definition.
• Add intermediate-level grouping functions to generate a point nowcast from a reporting triangle and to take a point nowcast and estimate and apply uncertainty.
• Refactor estimate_uncertainty to take in an error model function, an aggregator function for aggregating across reference times, and an aggregator function for aggregating across delays.

Documentation

• Add a vignette which walks through the low-level function interface on the same nowcasting problem as in the Getting Started vignette.
• Modify the Getting Started vignette to use the baselinenowcast() wrapper function.
• Improve PMF validation message to be more informative when the delay PMF does not sum to approximately 1. The message now shows the actual sum and clarifies that this may be expected when working with downward corrections or incomplete data (#148, #278).
• Include the pre-print as well as the package in the citation.

baselinenowcast 0.0.0.1.000

Package

• Replace argument names with more complete versions where possible.
• Replace most function names with action-oriented function naming.
• Add a check to ensure that there are sufficient non-zero values in the reporting triangle.
• Add a check to ensure that sufficient n are specified for the delay estimate.
• Change the requirement so that the number of rows used for delay estimation need not be greater than or equal to the number of columns, but instead that at least one row contains a full set of observations.
• Add support for passing in a restricted set of functions to the estimate_dispersion() function to transform the “target” dataset across reference dates.
• Implements a safe iterator in the step where retrospective point nowcasts are generated from a list of retrospective reporting triangles, ensuring that the iterations continue even if not all retrospective point nowcasts can be generated.
• Replace function named replace_lower_right_with_NA() with generate_triangle().
• Removes requirement that all elements of the reporting triangle are integers.
• Modify the method used to estimate dispersion. Instead of estimating dispersion on individual elements of the nowcast matrix, we now estimate dispersion on the predicted nowcasts summed across reporting dates as a function of forecast horizon.
• Standardises naming of objects that are matrices vs vectors and objects that contain observations, point estimates, and probabilistic draws.
• Modifies functions that estimate a delay and generate a point nowcast to ensure that they throw an error/warning if the first element of the delay PMF is 0.
• Adjusts function to estimate delay distribution to be able to handle complete and partially complete reporting triangles.
• Add function to convert a list of expected observed reporting squares to a long tidy dataframe indexed by reference time, delay, and a sample index.
• Implement zero-handling in the bottom left of the reporting triangle when applying the delay to generate a point nowcast.
• Add function to generate a list of expected observed reporting squares.
• Add function to generate an expected observed reporting square from a point nowcast and a vector of dispersion parameters.
• Add function to estimate dispersion parameters from a match list of nowcasts and observed reporting triangles.
• Add functions to generate retrospective nowcasts from a single reporting triangle.
• Refactor uncertainty estimation to use a user-facing function to generate retrospective reporting triangles.
• Introduced function to estimate the uncertainty from a triangle to be nowcasted and a delay distribution.
• Introduced functions to get the delay estimate and apply the delay, and used these in the Getting Started vignette.
• Added package skeleton.

Documentation

• Modify vignette to be consistent with the decided upon defaults for the number of reference times used for delay estimation and uncertainty.
• Add new logo.
• Methods write-up as a separate vignette.

Bug fixes

• Fix internal checks that ensure there is sufficient data for the specified target choice, using the number of NA rows rather than the number of columns as a proxy for the number of horizons.
• Bug fix to change the requirement so that the sum of the elements in the structure vector must not be greater than the number of columns.
• Introduced function to estimate the uncertainty from a triangle to be nowcasted and a delay distribution.
• Introduced functions to get the delay estimate and apply the delay, and used these in the Getting Started vignette.
• Added package skeleton.