forecast ILI

[stephenturner]

First cut at function to forecast ILI (fixes #7 when closed).

Example usage below. Object of interest in the returned list is ili_bound, which gives you a regular (non-tsibble) tibble with historical recorded data and horizon number of forecasted weeks into the future, bound in a tibble.

library(fiphde)

# Get data
ilidat <- get_cdc_ili(region="national", years=2010:lubridate::year(lubridate::today()))
#> Latest week_start / year / epiweek available:
#> 2021-12-05 / 2021 / 49

# Using data only from march 2020 forward
ilifor_2020 <- forecast_ili(ilidat, horizon=4L, location="US", trim_date="2020-03-01")
head(ilifor_2020$ili_bound)
#> # A tibble: 6 × 5
#>   location  year  week weighted_ili forecasted
#>   <chr>    <dbl> <dbl>        <dbl> <lgl>     
#> 1 US        2020    11         5.67 FALSE     
#> 2 US        2020    12         6.38 FALSE     
#> 3 US        2020    13         4.92 FALSE     
#> 4 US        2020    14         3.79 FALSE     
#> 5 US        2020    15         2.80 FALSE     
#> 6 US        2020    16         2.13 FALSE
tail(ilifor_2020$ili_bound)
#> # A tibble: 6 × 5
#>   location  year  week weighted_ili forecasted
#>   <chr>    <dbl> <dbl>        <dbl> <lgl>     
#> 1 US        2021    48         2.45 FALSE     
#> 2 US        2021    49         2.68 FALSE     
#> 3 US        2021    50         2.67 TRUE      
#> 4 US        2021    51         2.85 TRUE      
#> 5 US        2021    52         2.86 TRUE      
#> 6 US        2022     1         2.87 TRUE
ilifor_2020$ili_fit
#> # A mable: 1 x 2
#> # Key:     location [1]
#>   location          arima
#>   <chr>           <model>
#> 1 US       <ARIMA(0,2,3)>
ilifor_2020$ili_fit %>% focustools::extract_arima_params()
#> # A tibble: 3 × 22
#>   location abbreviation location_name .model     p     d     q     P     D     Q
#>   <chr>    <chr>        <chr>         <chr>  <int> <int> <int> <dbl> <dbl> <dbl>
#> 1 US       US           US            arima      0     2     3     0     0     0
#> 2 US       US           US            arima      0     2     3     0     0     0
#> 3 US       US           US            arima      0     2     3     0     0     0
#> # … with 12 more variables: constant <lgl>, period <dbl>, term <chr>,
#> #   estimate <dbl>, std.error <dbl>, statistic <dbl>, p.value <dbl>,
#> #   sigma2 <dbl>, log_lik <dbl>, AIC <dbl>, AICc <dbl>, BIC <dbl>
ilifor_2020$ili_forecast
#> # A fable: 4 x 5 [1W]
#> # Key:     location, .model [1]
#>   location .model    yweek  weighted_ili .mean
#>   <chr>    <chr>    <week>        <dist> <dbl>
#> 1 US       arima  2021 W50 N(2.7, 0.048)  2.67
#> 2 US       arima  2021 W51  N(2.9, 0.16)  2.85
#> 3 US       arima  2021 W52  N(2.9, 0.43)  2.86
#> 4 US       arima  2022 W01   N(2.9, 0.7)  2.87

# Using all the data we have (2010-forward, in this example)
ilifor_2010 <- forecast_ili(ilidat, horizon=4L, location="US")
head(ilifor_2010$ili_bound)
#> # A tibble: 6 × 5
#>   location  year  week weighted_ili forecasted
#>   <chr>    <dbl> <dbl>        <dbl> <lgl>     
#> 1 US        2010    40         1.11 FALSE     
#> 2 US        2010    41         1.24 FALSE     
#> 3 US        2010    42         1.26 FALSE     
#> 4 US        2010    43         1.26 FALSE     
#> 5 US        2010    44         1.43 FALSE     
#> 6 US        2010    45         1.52 FALSE
tail(ilifor_2010$ili_bound)
#> # A tibble: 6 × 5
#>   location  year  week weighted_ili forecasted
#>   <chr>    <dbl> <dbl>        <dbl> <lgl>     
#> 1 US        2021    48         2.45 FALSE     
#> 2 US        2021    49         2.68 FALSE     
#> 3 US        2021    50         2.77 TRUE      
#> 4 US        2021    51         2.76 TRUE      
#> 5 US        2021    52         2.72 TRUE      
#> 6 US        2022     1         2.69 TRUE

See help for ?fiphde::forecast_ili.

Currently only tested with location="US". Possible we could use location as a key and retrieve forecasted ILI for all states. Keeping this in draft until I get something tested there.

[vpnagraj]

@stephenturner fyi i just ported over the glm functions code from fluforce-init

also moved the example code from fluforce-init into scratch (saw you mention something similar in a commit on this PR so wanted you to have a heads up):

https://github.com/signaturescience/fiphde/blob/main/scratch/worfklow.R

all seems to be working by using glm functions now in fiphde package

[stephenturner]

Nice, thanks for that. I'm running through the old code in fluforce-init and it's taking forever. I'm going to merge in main on this branch but it looks like we have a merge conflict I'll work out

[stephenturner]

I updated workflow.R to use the package functions for forecasting ILI, and replaced the hard-coded values.

[stephenturner]

Brought in state-level forecasting in #31. Now using unweighted ILI, and removes states with too much missing data (now, only FL). When merged this will fix #27

[vpnagraj]

issue with the workflow.R script should be fixed now in 8d85d50

[vpnagraj]

oh and note column renaming (year=>epiyear, week=>epiweek ) in 8d85d50

[stephenturner]

I'm a little unhappy with the ili forecast's sensitivity to trim dates. Left constrained, right unconstrained, different dates in rows.

[vpnagraj]

still some rough edges to polish with docs / code cleanup. but overall this looks good enough to merge to main to me. i did make one change to add a "param_space" named list arg to forecast_ili(). so when you have a chance @stephenturner take a look and confirm that is functional on your end. then lets go ahead and merge this branch to main.

[stephenturner]

I think this works after a small change in e11a52c. One thing to note, you must specify all 6 params even if you only want to restrict one. But works, and rcmdchecks OK