heree is a potentialpacage [originally proposed by @benjaminortizulloa]

Google has started providing mobility statistics derived from mobility data from smartphone users: https://www.google.com/covid19/mobility/.

Description

For public health officials and researchers it will be crucial to be able to follow the evolution of those mobility indicators through time and to integrate them in automated analyses. A means to automatically access those indicators (and their evolution) from within R code is thus needed.

1. See if google provides an API to automatically access indicators
   • If yes: go to 2.
   • If no, examine pdf reports to see if indicators can be extracted automatically
2. provide an R package that makes the indicators available from within R in tabular form through API or by extracting values from pdf reports

Output

The proposed output is a data frame in long format with columns for country, date, indicator and values.

Impact

This will allow to follow mobility indicators automatically from within R code, which will allow for analysis of (for example) the impact of mobility reduction on transmissibility.

Proposed Timeline

First version of package available on Apr 10.

Focal Point

@ffinger

Links

Could be integrated into https://github.com/epiforecasts/NCoVUtils or live as a separate package.

Package: fit discretised distributions to match expected distributional properties

Description

Disclaimer: if this exists already, please feel free to point it out here.

What we need: fit a discretised distribution (distcrete object) so that it matches user-specified properties e.g. quantiles (typically the median, and 25% / 75% quantiles)

The package should include documentation, simple reproducible examples, unit tests, and use good practices and standards recommended by RECON or similar.

Specs (option 1)

This version would be slightly easier to use, but more constrained as we explicitely force fitting using quantiles. The method for calculating distances from targets is likely pre-programmed in the function, so only a few options will be available (likely: least square, absolute values, etc).

• inputs:
  • name: a single character indicating the family of a distribution; I would recommend using gamma and weibull for testing purposes, as these are often the distributions we end up using; see name in ?distcrete::distcrete
  • target_quantiles: a numeric vector providing quantile used for fitting (between 0 and 1)
  • target_values: a numeric vector providing quantiles values to be fitted to
  • fit_method: a single character indicating which fitting method should be used (default to least squares)
  • ... other arguments passed to distcrete::distcrete
• outputs:
  • a distcrete object with quantile properties matching the user-specified quantiles as closely as possible

Specs (option 2)

This version would be more flexible targets and fitting are passed as user-specified functions.

• inputs:
  • name: a single character indicating the family of a distribution; I would recommend using gamma and weibull for testing purposes, as these are often the distributions we end up using; see name in ?distcrete::distcrete
  • summary: a function with a single argument x calculating a distributional summary of its input
  • distance: a function with 2 arguments x and y, calculating a dissimilarity between x and y, as a single numeric, finite value
  • ... other arguments passed to distcrete::distcrete
• outputs:
  • a distcrete object with quantile properties matching the user-specified quantiles as closely as possible

Relevant related packages

• distcrete

Starting point

The following code generates a discretised weibull as a distcrete object, generates a sample of 1e5 values, and calculates the sum of squared distances from reference quantiles:

## define targets
target_quantiles <- c(.25, .5, .75)
target_values <- c(3, 5, 7)


## make distcrete object
x <- distcrete::distcrete("weibull", shape = 2, scale = 4, w = 0.5, interval = 1)

## check quantiles
x_quantiles <- quantile(x$r(1e5), probs = target_quantiles)
sum_square <- sum((x_quantiles - target_values)^2)

Further tweaks will be putting this insided an optimiser. The interface should allow accessing all arguments of distcrete::distcrete in addition to arguments used to specify targets and fitting

Impact

Many epidemiologically relevant delays are shared in the literature as means, medians, IQR, 95%CI etc. However, it can be very tricky to turn these into discretised distributions which we can then use for forecasting, estimating transmissibility etc. A simple utilitary function implementing this would be a great help for this.

Proposed timeline

• first viable product: as soon as possible; proposed: 25th April
• first stable version (CRAN release): less urgent, the main thing is to have this package, tested, documented, and publicly accessible

Focal point

@thibautjombart

Description

A function for each country that accesses public data sources and makes the data accessible in R.
Good examples in this package: https://github.com/epiforecasts/NCoVUtils for a number of countries.

Functions are existent for most European and Asian countries and the US. We are looking for data and functions for LMIC at the moment, especially African countries.

I suggest we use this issue to keep track of

1. Needs for data and R functions for specific countries
2. Available data sources for specific countries
3. Functions that import the data sources identified in 2.

See below for google sheet tracking those.

Output

The output format of each function should be a long data frame containing the following columns:

• country
• admin_subdivision_level_1
• admin_subdivision_level_2 (if available)
• (more levels if available)
• date
• cases
• deaths

where cases and deaths stand for the newly reported cases and deaths on that day.

Functions can either be added to https://github.com/epiforecasts/NCoVUtils via pull request, or we can start our own package that wraps NCoVUtils and other solutions for the already implemented countries.

Countries already covered:

https://github.com/epiforecasts/NCoVUtils covers the following countries so far:

• Belgium
• Canada
• France
• Germany
• Italy
• Spain
• United Kingdom
• United States
• Japan
• Korea
• Afghanistan

Countries to be done

Spreadsheet to track requested countries, data sources and implementations:
https://docs.google.com/spreadsheets/d/1uvg07BAmwKqLqhKvkejhkX7uvXiGCre4sz11Au3pz9Q/edit?usp=sharing

my own priority list:

• Burkina Faso
• Irak
• Democratic Republic of the Congo
• Syria

Links

A few places where data sources are indexed:

https://data.humdata.org/event/covid-19
https://coronavirustechhandbook.com/home
https://www.europeandataportal.eu/data/datasets?locale=en&categories=heal&page=1&query=covid

Package: bed occupancy forecasting

Description

Package existing code to implement forecasting of daily bed occupancy from input on daily admissions and the distribution of duration of stay.

The package should include documentation, simple reproducible examples, unit tests.

Use cases / specs

1. predict bed occupancy from existing admission data
   • input dates of admissions as Date
   • input numbers of admissions as integer
   • input distribution of the length of stay (LoS) as a distcrete object (preferably) or as a function returning integer numbers along the lines of rgeom()
2. do the above, but with several admission trajectories (e.g. from different models)
   • input on admissions as a projections object
   • input LoS as in 1.
3. outputs should be projections objects

Relevant related pacakges

• projections; note the current github version should be used
• distcrete

Starting point

A starting point is provided in the two following gists:

• simulate_occupancy.R
• project_beds.R

Impact

Some of the main questions for COVID19 response relate to hospital bed capacity. For this, multpile modelling groups are making forecastings of bed occupancy, e.g. our app. However there is no consistent way of doing this, leaving room for potential errors and consistency across results. Having a reliable, computer-efficient, well documented and tested way of forecasting bed occupancy will improve the quality of, and the ease for providing, these predictions.

Proposed timeline

• first viable product: as soon as possible; end of May?
• first stable version (CRAN release): less urgent, the main thing is to have this package, tested, documented, and publicly accessible

Focal point

@thibautjombart

Help develop a ... (title of the issue, potentially longer)

*This template is provided to help structure tasks. Feel free to elaborate if you feel like the provided structure is not sufficient to describe your task. Also, please feel free to provide feedback on the template by creating a separate issue and using the label internal.

Description

This is some information on the task. Keep it short, and don't hesitate to use bullet/numbered lists like:

1. first we need this
2. then we need that
3. ...
   Or use tick boxes:

• first we need this
• then we need that
• ...

Impact

This project will be used for ... and improve...

Proposed timeline

• first viable product: [a date]
• first stable version: [a date]

Focal point

Tag the person (if not yourself) to contact for this project using their github ID.

Links

Provide some links here if relevant, including other github issues that may alread provide information on the task.

information about the repo [originally proposed by @benjaminortizulloa]

test [originally proposed by @benjaminortizulloa]

Make apple mobility data available from R.
https://www.apple.com/covid19/mobility

It's already a CSV, so very easy to wrap it.

Want to include this in your package @nacnudus?
Or in NCoVUtils @seabbs?

NCoVUtils is a package to access a number of national and international data sources in a consistent way.
Cross-posting their Issues page here, help needed updating outdated data sources, adding new, or improve the package.

Please comment directly in the relevant issue.

https://github.com/epiforecasts/NCoVUtils/issues

this is a test task [originally proposed by @matthieugold]